Deep Reinforcement Learning – Winter 2021/22

In recent years, reinforcement learning has been combined with deep neural networks, giving rise to game agents with super-human performance (for example for Go, chess, or 1v1 Dota2, capable of being trained solely by self-play), datacenter cooling algorithms being 50% more efficient than trained human operators, or improved machine translation. The goal of the course is to introduce reinforcement learning employing deep neural networks, focusing both on the theory and on practical implementations.

Python programming skills and TensorFlow skills (or any other deep learning framework) are required, to the extent of the NPFL114 course. No previous knowledge of reinforcement learning is necessary.

About

SIS code: NPFL122
Semester: winter
E-credits: 5
Examination: 2/2 C+Ex
Guarantor: Milan Straka

Timespace Coordinates

  • lecture: the lecture is held on Monday 9:00 in S5; first lecture is on Oct 04
  • practicals: the practicals take place on Monday 10:40 in S5; first practicals are on Oct 04

Lectures

1. Introduction to Reinforcement Learning Slides PDF Slides Lecture Questions bandits monte_carlo

2. Markov Decision Process, Optimal Solutions, Monte Carlo Methods Slides PDF Slides Lecture Questions policy_iteration policy_iteration_exact policy_iteration_exploring_mc policy_iteration_greedy_mc

3. Temporal Difference Methods, Off-Policy Methods Slides PDF Slides Lecture Questions importance_sampling q_learning lunar_lander td_algorithms

4. Function Approximation, Deep Q Network Slides PDF Slides Lecture Questions q_learning_tiles q_network

5. Rainbow Slides PDF Slides Lecture Questions car_racing

6. Policy Gradient Methods Slides PDF Slides Lecture Questions reinforce reinforce_baseline cart_pole_pixels

7. Continuous Actions, DDPG, TD3 Slides PDF Slides Lecture Questions paac paac_continuous ddpg

8. SAC, Eligibility Traces Slides PDF Slides Lecture Questions walker walker_hardcore

9. Eligibility traces, Impala, R2D2, Agent57 Slides PDF Slides Lecture Questions brax_cheetah trace_algorithms

10. UCB, Monte Carlo Tree Search, AlphaZero Slides PDF Slides Lecture Questions az_quiz az_quiz_randomized

11. MuZero, PlaNet Slides PDF Slides Lecture Questions

12. ST and Gumbel-Softmax, DreamerV2, MERLIN, FTW Slides PDF Slides Lecture Questions memory_game memory_game_rl

13. Multi-Agent RL, PPO, MAPPO Slides PDF Slides Lecture ppo mappo

The lecture content, including references to study materials.

The main study material is the Reinforcement Learning: An Introduction; second edition by Richard S. Sutton and Andrew G. Barto (reffered to as RLB). It is available online and also as a hardcopy.

References to study materials cover all theory required at the exam, and sometimes even more – the references in italics cover topics not required for the exam.

1. Introduction to Reinforcement Learning

 Oct 04 Slides PDF Slides Lecture Questions bandits monte_carlo

  • History of RL [Chapter 1 of RLB]
  • Multi-armed bandits [Sections 2-2.6 of RLB]
  • Markov Decision Process [Sections 3-3.3 of RLB]

2. Markov Decision Process, Optimal Solutions, Monte Carlo Methods

 Oct 11 Slides PDF Slides Lecture Questions policy_iteration policy_iteration_exact policy_iteration_exploring_mc policy_iteration_greedy_mc

  • Policies and Value Functions [Sections 3.5-3.6 of RLB]
  • Value Iteration [Sections 4 and 4.4 of RLB]
    • Proof of convergence only in slides
  • Policy Iteration [Sections 4.1-4.3 of RLB]
  • Generalized Policy Iteration [Section 4.6 or RLB]
  • Monte Carlo Methods [Sections 5-5.4 of RLB]

3. Temporal Difference Methods, Off-Policy Methods

 Oct 18 Slides PDF Slides Lecture Questions importance_sampling q_learning lunar_lander td_algorithms

  • Model-free and model-based methods, using state-value or action-value functions [Chapter 8 before Section 8.1, and Section 6.8 of RLB]
  • Temporal-difference methods [Sections 6-6.3 of RLB]
  • Sarsa [Section 6.4 of RLB]
  • Q-learning [Section 6.5 of RLB]
  • Off-policy Monte Carlo Methods [Sections 5.5-5.7 of RLB]
  • Expected Sarsa [Section 6.6 of RLB]
  • N-step TD policy evaluation [Section 7.1 of RLB]
  • N-step Sarsa [Section 7.2 of RLB]
  • Off-policy n-step Sarsa [Section 7.3 of RLB]
  • Tree backup algorithm [Section 7.5 of RLB]

4. Function Approximation, Deep Q Network

 Oct 25 Slides PDF Slides Lecture Questions q_learning_tiles q_network

  • Function approximation [Sections 9-9.3 of RLB]
  • Tile coding [Section 9.5.4 of RLB]
  • Linear function approximation [Section 9.4 of RLB, without the Proof of Convergence if Linear TD(0)]
  • Semi-Gradient TD methods [Sections 9.3, 10-10.2 of RLB]
  • Off-policy function approximation TD divergence [Sections 11.2-11.3 of RLB]
  • Deep Q Network [Volodymyr Mnih et al.: Human-level control through deep reinforcement learning]

5. Rainbow

 Nov 01 Slides PDF Slides Lecture Questions car_racing

6. Policy Gradient Methods

 Nov 08 Slides PDF Slides Lecture Questions reinforce reinforce_baseline cart_pole_pixels

7. Continuous Actions, DDPG, TD3

 Nov 15 Slides PDF Slides Lecture Questions paac paac_continuous ddpg

8. SAC, Eligibility Traces

 Nov 22 Slides PDF Slides Lecture Questions walker walker_hardcore

9. Eligibility traces, Impala, R2D2, Agent57

 Nov 29 Slides PDF Slides Lecture Questions brax_cheetah trace_algorithms

10. UCB, Monte Carlo Tree Search, AlphaZero

 Dec 06 Slides PDF Slides Lecture Questions az_quiz az_quiz_randomized

11. MuZero, PlaNet

 Dec 13 Slides PDF Slides Lecture Questions

12. ST and Gumbel-Softmax, DreamerV2, MERLIN, FTW

 Dec 20 Slides PDF Slides Lecture Questions memory_game memory_game_rl

13. Multi-Agent RL, PPO, MAPPO

 Jan 03 Slides PDF Slides Lecture ppo mappo

Requirements

To pass the practicals, you need to obtain at least 80 points, excluding the bonus points. Note that all surplus points (both bonus and non-bonus) will be transfered to the exam. In total, assignments for at least 120 points (not including the bonus points) will be available, and if you solve all the assignments (either before the first or the second deadline), you automatically pass the exam with grade 1.

Environment

The tasks are evaluated automatically using the ReCodEx Code Examiner.

The evaluation is performed using Python 3.9, TensorFlow 2.6.0, TensorFlow Probability 0.14.1, NumPy 1.19.5 and OpenAI Gym 0.20.0. You should install the exact version of these packages yourselves. For those using PyTorch, 1.9.1 is also available.

Teamwork

Solving assignments in teams (of size at most 3) is encouraged, but everyone has to participate (it is forbidden not to work on an assignment and then submit a solution created by other team members). All members of the team must submit in ReCodEx individually, but can have exactly the same sources/models/results. Each such solution must explicitly list all members of the team to allow plagiarism detection using this template.

No Cheating

Cheating is strictly prohibited and any student found cheating will be punished. The punishment can involve failing the whole course, or, in grave cases, being expelled from the faculty. While discussing assignments with any classmate is fine, each team must complete the assignments themselves, without using code they did not write (unless explicitly allowed). Of course, inside a team you are expected to share code and submit identical solutions.

bandits

 Deadline: Oct 17, 23:59  3 points

Implement the εε-greedy strategy for solving multi-armed bandits.

Start with the bandits.py template, which defines MultiArmedBandits environment, which has the following two methods:

  • reset(): reset the environment
  • step(action) → reward: perform the chosen action in the environment, obtaining a reward
  • greedy(epsilon): return True with probability 1-epsilon

Your goal is to implement the following solution variants:

  • alpha=0=0: perform εε-greedy search, updating the estimated using averaging.
  • alpha0≠0: perform εε-greedy search, updating the estimated using a fixed learning rate alpha.

Note that the initial estimates should be set to a given value and epsilon can be zero, in which case purely greedy actions are used.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 bandits.py --alpha=0 --epsilon=0.1 --initial=0
1.39 0.08
  • python3 bandits.py --alpha=0 --epsilon=0 --initial=1
1.48 0.22
  • python3 bandits.py --alpha=0.15 --epsilon=0.1 --initial=0
1.37 0.09
  • python3 bandits.py --alpha=0.15 --epsilon=0 --initial=1
1.52 0.04

monte_carlo

 Deadline: Oct 17, 23:59  5 points

Solve the discretized CartPole-v1 environment environment from the OpenAI Gym using the Monte Carlo reinforcement learning algorithm. The gym environments have the followng methods and properties:

  • observation_space: the description of environment observations
  • action_space: the description of environment actions
  • reset() → new_state: starts a new episode
  • step(action) → new_state, reward, done, info: perform the chosen action in the environment, returning the new state, obtained reward, a boolean flag indicating an end of episode, and additional environment-specific information
  • render(): render current environment state

We additionaly extend the gym environment by:

  • episode: number of the current episode (zero-based)
  • reset(start_evaluation=False) → new_state: if start_evaluation is True, an evaluation is started

Once you finish training (which you indicate by passing start_evaluation=True to reset), your goal is to reach an average return of 490 during 100 evaluation episodes. Note that the environment prints your 100-episode average return each 10 episodes even during training.

Start with the monte_carlo.py template, which parses several useful parameters, creates the environment and illustrates the overall usage.

During evaluation in ReCodEx, three different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 5 minutes.

policy_iteration

 Deadline: Oct 24, 23:59  2 points

Consider the following gridworld:

Gridworld example

Start with policy_iteration.py, which implements the gridworld mechanics, by providing the following methods:

  • GridWorld.states: return number of states (11)
  • GridWorld.actions: return lists with labels of the actions (["↑", "→", "↓", "←"])
  • GridWorld.step(state, action): return possible outcomes of performing the action in a given state, as a list of triples containing
    • probability: probability of the outcome
    • reward: reward of the outcome
    • new_state: new state of the outcome

Implement policy iteration algorithm, with --steps steps of policy evaluation/policy improvement. During policy evaluation, use the current value function and perform --iterations applications of the Bellman equation. Perform the policy evaluation asynchronously (i.e., update the value function in-place for states 0,1,0, 1, …). Assume the initial policy is “go North” and initial value function is zero.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 policy_iteration.py --gamma=0.95 --iterations=1 --steps=1
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑           -10.00←  -10.95↑
    0.00↑    0.00←   -7.50←  -88.93←
  • python3 policy_iteration.py --gamma=0.95 --iterations=1 --steps=2
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑            -8.31←  -11.83←
    0.00↑    0.00←   -1.50←  -20.61←
  • python3 policy_iteration.py --gamma=0.95 --iterations=1 --steps=3
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑            -6.46←   -6.77←
    0.00↑    0.00←   -0.76←  -13.08↓
  • python3 policy_iteration.py --gamma=0.95 --iterations=1 --steps=10
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑            -1.04←   -0.83←
    0.00↑    0.00←   -0.11→   -0.34↓
  • python3 policy_iteration.py --gamma=0.95 --iterations=10 --steps=10
   11.93↓   11.19←   10.47←    6.71↑
   12.83↓            10.30←   10.12←
   13.70→   14.73→   15.72→   16.40↓
  • python3 policy_iteration.py --gamma=1 --iterations=1 --steps=100
   74.73↓   74.50←   74.09←   65.95↑
   75.89↓            72.63←   72.72←
   77.02→   78.18→   79.31→   80.16↓

policy_iteration_exact

 Deadline: Oct 24, 23:59  2 points

Starting with policy_iteration_exact.py, extend the policy_iteration assignment to perform policy evaluation exactly by solving a system of linear equations.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 policy_iteration_exact.py --gamma=0.95 --steps=1
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑           -12.35←  -12.35↑
   -0.85←   -8.10←  -19.62← -100.71←
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=2
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑             0.00←  -11.05←
    0.00↑    0.00↑    0.00←  -12.10↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=3
   -0.00↑   -0.00↑   -0.00↑   -0.00↑
   -0.00↑            -0.00←    0.69←
   -0.00↑   -0.00↑   -0.00→    6.21↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=4
    0.00↑    0.00↑    0.00↓    0.00↑
    0.00↓             5.91←    6.11←
    0.65→    6.17→   14.93→   15.99↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=5
    2.83↓    4.32→    8.09↓    5.30↑
   12.92↓             9.44←    9.35←
   13.77→   14.78→   15.76→   16.53↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=6
   11.75↓    8.15←    8.69↓    5.69↑
   12.97↓             9.70←    9.59←
   13.82→   14.84→   15.82→   16.57↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=7
   12.12↓   11.37←    9.19←    6.02↑
   13.01↓             9.92←    9.79←
   13.87→   14.89→   15.87→   16.60↓
  • python3 policy_iteration_exact.py --gamma=0.95 --steps=8
   12.24↓   11.49←   10.76←    7.05↑
   13.14↓            10.60←   10.42←
   14.01→   15.04→   16.03→   16.71↓

policy_iteration_exploring_mc

 Deadline: Oct 24, 23:59  2 points

Starting with policy_iteration_exploring_mc.py, extend the policy_iteration assignment to perform policy evaluation by using Monte Carlo estimation with exploring starts. Specifically, we update the action-value function qπ(s,a)q_\pi(s, a) by running a simulation with a given number of steps and using the observed return as its estimate.

The estimation can now be performed model-free (without the access to the full MDP dynamics), therefore, the GridWorld.step returns a randomly sampled result instead of a full distribution.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 policy_iteration_exploring_mc.py --gamma=0.95 --seed=42 --steps=1
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑             0.00↑    0.00↑
    0.00↑    0.00→    0.00↑    0.00↓
  • python3 policy_iteration_exploring_mc.py --gamma=0.95 --seed=42 --steps=10
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑             0.00↑  -19.50↑
    0.27↓    0.48←    2.21↓    8.52↓
  • python3 policy_iteration_exploring_mc.py --gamma=0.95 --seed=42 --steps=50
    0.09↓    0.32↓    0.22←    0.15↑
    0.18↑            -2.43←   -5.12↓
    0.18↓    1.80↓    3.90↓    9.14↓
  • python3 policy_iteration_exploring_mc.py --gamma=0.95 --seed=42 --steps=100
    3.09↓    2.42←    2.39←    1.17↑
    3.74↓             1.66←    0.18←
    3.92→    5.28→    7.16→   11.07↓
  • python3 policy_iteration_exploring_mc.py --gamma=0.95 --seed=42 --steps=200
    7.71↓    6.76←    6.66←    3.92↑
    8.27↓             6.17←    5.31←
    8.88→   10.12→   11.36→   13.92↓

policy_iteration_greedy_mc

 Deadline: Oct 24, 23:59  2 points

Starting with policy_iteration_greedy_mc.py, extend the policy_iteration_exploring_mc assignment to perform policy evaluation by using εε-greedy Monte Carlo estimation. Specifically, we update the action-value function qπ(s,a)q_\pi(s, a) by running a simulation with a given number of steps and using the observed return as its estimate.

For the sake of replicability, use the provided GridWorld.epsilon_greedy(epsilon, greedy_action) method, which returns a random action with probability of epsilon and otherwise returns the given greedy_action.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=1
    0.00↑    0.00↑    0.00↑    0.00↑
    0.00↑             0.00→    0.00→
    0.00↑    0.00↑    0.00→    0.00→
  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=10
   -1.20↓   -1.43←    0.00←   -6.00↑
    0.78→           -20.26↓    0.00←
    0.09←    0.00↓   -9.80↓   10.37↓
  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=50
   -0.16↓   -0.19←    0.56←   -6.30↑
    0.13→            -6.99↓   -3.51↓
    0.01←    0.00←    3.18↓    7.57↓
  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=100
   -0.07↓   -0.09←    0.28←   -4.66↑
    0.06→            -5.04↓   -8.32↓
    0.00←    0.00←    1.70↓    4.38↓
  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=200
   -0.04↓   -0.04←   -0.76←   -4.15↑
    0.03→            -8.02↓   -5.96↓
    0.00←    0.00←    2.53↓    4.36↓
  • python3 policy_iteration_greedy_mc.py --gamma=0.95 --seed=42 --steps=500
   -0.02↓   -0.02←   -0.65←   -3.52↑
    0.01→           -11.34↓   -8.07↓
    0.00←    0.00←    3.15↓    3.99↓

importance_sampling

 Deadline: Oct 31, 23:59  2 points

Using the FrozenLake-v1 environment environment, implement Monte Carlo weighted importance sampling to estimate state value function VV of target policy, which uniformly chooses either action 1 (down) or action 2 (right), utilizing behaviour policy, which uniformly chooses among all four actions.

Start with the importance_sampling.py template, which creates the environment and generates episodes according to behaviour policy.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 importance_sampling.py --episodes=500
 0.00  0.00  0.00  0.00
 0.03  0.00  0.00  0.00
 0.22  0.14  0.29  0.00
 0.00  0.50  1.00  0.00
  • python3 importance_sampling.py --episodes=5000
 0.00  0.01  0.02  0.00
 0.00  0.00  0.08  0.00
 0.06  0.08  0.17  0.00
 0.00  0.19  0.89  0.00
  • python3 importance_sampling.py --episodes=50000
 0.02  0.01  0.04  0.01
 0.03  0.00  0.06  0.00
 0.08  0.17  0.24  0.00
 0.00  0.34  0.78  0.00

q_learning

 Deadline: Oct 31, 23:59  4 points

Solve the MountainCar-v0 environment environment from the OpenAI Gym using the Q-learning reinforcement learning algorithm. Note that this task does not require TensorFlow.

The environment methods and properties are described in the monte_carlo assignment. Once you finish training (which you indicate by passing start_evaluation=True to reset), your goal is to reach an average return of -150 during 100 evaluation episodes.

You can start with the q_learning.py template, which parses several useful parameters, creates the environment and illustrates the overall usage. Note that setting hyperparameters of Q-learning is a bit tricky – I usualy start with a larger value of εε (like 0.2 or even 0.5) an then gradually decrease it to almost zero.

During evaluation in ReCodEx, three different random seeds will be employed, and you need to reach the required return on all of them. The time limit for each test is 5 minutes.

lunar_lander

 Deadline: Oct 31, 23:59  7 points + 7 bonus

Solve the LunarLander-v2 environment environment from the OpenAI Gym. Note that this task does not require TensorFlow.

The environment methods and properties are described in the monte_carlo assignment, but include one additional method:

  • expert_trajectory() → initial_state, trajectory This method generates one expert trajectory and returns a pair of initial_state and trajectory, where trajectory is a list of the tripples (action, reward, next_state). You can use this method only during training, not during evaluation.

To pass the task, you need to reach an average return of 0 during 1000 evaluation episodes. During evaluation in ReCodEx, three different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 15 minutes.

The task is additionally a competition and at most 7 points will be awarded according to relative ordering of your solution performances.

You can start with the lunar_lander.py template, which parses several useful parameters, creates the environment and illustrates the overall usage.

td_algorithms

 Deadline: Nov 07, 23:59  4 points

Starting with the td_algorithms.py template, implement all of the following nn-step TD methods variants:

  • SARSA, expected SARSA and Tree backup;
  • either on-policy (with εε-greedy behaviour policy) or off-policy (with the same behaviour policy, but greedy target policy).

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 td_algorithms.py --mode=sarsa --n=1
Episode 200, mean 100-episode return -223.71 +-93.59
Episode 400, mean 100-episode return -139.87 +-94.98
Episode 600, mean 100-episode return -79.92 +-78.13
Episode 800, mean 100-episode return -45.47 +-61.87
Episode 1000, mean 100-episode return -23.00 +-37.69
  • python3 td_algorithms.py --mode=sarsa --n=1 --off_policy
Episode 200, mean 100-episode return -226.79 +-86.95
Episode 400, mean 100-episode return -121.31 +-98.15
Episode 600, mean 100-episode return -75.73 +-82.94
Episode 800, mean 100-episode return -32.20 +-54.12
Episode 1000, mean 100-episode return -12.77 +-31.50
  • python3 td_algorithms.py --mode=sarsa --n=4
Episode 200, mean 100-episode return -250.47 +-162.91
Episode 400, mean 100-episode return -84.50 +-136.47
Episode 600, mean 100-episode return -13.44 +-46.17
Episode 800, mean 100-episode return -2.66 +-14.03
Episode 1000, mean 100-episode return -1.50 +-12.57
  • python3 td_algorithms.py --mode=sarsa --n=4 --off_policy
Episode 200, mean 100-episode return -355.38 +-114.57
Episode 400, mean 100-episode return -195.13 +-184.38
Episode 600, mean 100-episode return -35.85 +-89.91
Episode 800, mean 100-episode return -2.21 +-13.34
Episode 1000, mean 100-episode return -0.94 +-12.37
  • python3 td_algorithms.py --mode=expected_sarsa --n=1
Episode 200, mean 100-episode return -236.48 +-88.08
Episode 400, mean 100-episode return -130.54 +-97.12
Episode 600, mean 100-episode return -80.92 +-75.09
Episode 800, mean 100-episode return -45.32 +-58.35
Episode 1000, mean 100-episode return -18.23 +-33.90
  • python3 td_algorithms.py --mode=expected_sarsa --n=1 --off_policy
Episode 200, mean 100-episode return -211.49 +-96.62
Episode 400, mean 100-episode return -133.54 +-94.77
Episode 600, mean 100-episode return -61.63 +-71.41
Episode 800, mean 100-episode return -27.20 +-40.28
Episode 1000, mean 100-episode return -12.80 +-25.67
  • python3 td_algorithms.py --mode=expected_sarsa --n=4
Episode 200, mean 100-episode return -208.38 +-160.04
Episode 400, mean 100-episode return -60.27 +-117.63
Episode 600, mean 100-episode return -4.22 +-18.40
Episode 800, mean 100-episode return -1.46 +-11.79
Episode 1000, mean 100-episode return -1.50 +-17.14
  • python3 td_algorithms.py --mode=expected_sarsa --n=4 --off_policy
Episode 200, mean 100-episode return -360.77 +-127.22
Episode 400, mean 100-episode return -232.19 +-196.49
Episode 600, mean 100-episode return -71.42 +-142.10
Episode 800, mean 100-episode return -7.37 +-26.92
Episode 1000, mean 100-episode return -5.71 +-27.27
  • python3 td_algorithms.py --mode=tree_backup --n=1
Episode 200, mean 100-episode return -236.48 +-88.08
Episode 400, mean 100-episode return -130.54 +-97.12
Episode 600, mean 100-episode return -80.92 +-75.09
Episode 800, mean 100-episode return -45.32 +-58.35
Episode 1000, mean 100-episode return -18.23 +-33.90
  • python3 td_algorithms.py --mode=tree_backup --n=1 --off_policy
Episode 200, mean 100-episode return -211.49 +-96.62
Episode 400, mean 100-episode return -133.54 +-94.77
Episode 600, mean 100-episode return -61.63 +-71.41
Episode 800, mean 100-episode return -27.20 +-40.28
Episode 1000, mean 100-episode return -12.80 +-25.67
  • python3 td_algorithms.py --mode=tree_backup --n=4
Episode 200, mean 100-episode return -240.56 +-121.74
Episode 400, mean 100-episode return -67.22 +-120.22
Episode 600, mean 100-episode return -7.58 +-42.29
Episode 800, mean 100-episode return -7.12 +-41.08
Episode 1000, mean 100-episode return 0.61 +-8.93
  • python3 td_algorithms.py --mode=tree_backup --n=4 --off_policy
Episode 200, mean 100-episode return -302.59 +-137.42
Episode 400, mean 100-episode return -99.90 +-149.18
Episode 600, mean 100-episode return -9.82 +-36.69
Episode 800, mean 100-episode return -0.14 +-9.99
Episode 1000, mean 100-episode return 0.11 +-9.68

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 td_algorithms.py --episodes=10 --mode=sarsa --n=1
Episode 10, mean 100-episode return -650.00 +-103.40
  • python3 td_algorithms.py --episodes=10 --mode=sarsa --n=1 --off_policy
Episode 10, mean 100-episode return -575.50 +-161.10
  • python3 td_algorithms.py --episodes=10 --mode=sarsa --n=4
Episode 10, mean 100-episode return -736.10 +-63.01
  • python3 td_algorithms.py --episodes=10 --mode=sarsa --n=4 --off_policy
Episode 10, mean 100-episode return -602.20 +-174.02
  • python3 td_algorithms.py --episodes=10 --mode=expected_sarsa --n=1
Episode 10, mean 100-episode return -673.90 +-80.11
  • python3 td_algorithms.py --episodes=10 --mode=expected_sarsa --n=1 --off_policy
Episode 10, mean 100-episode return -632.00 +-109.27
  • python3 td_algorithms.py --episodes=10 --mode=expected_sarsa --n=4
Episode 10, mean 100-episode return -737.00 +-76.37
  • python3 td_algorithms.py --episodes=10 --mode=expected_sarsa --n=4 --off_policy
Episode 10, mean 100-episode return -560.90 +-147.25
  • python3 td_algorithms.py --episodes=10 --mode=tree_backup --n=1
Episode 10, mean 100-episode return -673.90 +-80.11
  • python3 td_algorithms.py --episodes=10 --mode=tree_backup --n=1 --off_policy
Episode 10, mean 100-episode return -632.00 +-109.27
  • python3 td_algorithms.py --episodes=10 --mode=tree_backup --n=4
Episode 10, mean 100-episode return -708.50 +-169.01
  • python3 td_algorithms.py --episodes=10 --mode=tree_backup --n=4 --off_policy
Episode 10, mean 100-episode return -695.00 +-97.85

q_learning_tiles

 Deadline: Nov 07, 23:59  3 points

Improve the q_learning task performance on the MountainCar-v0 environment environment using linear function approximation with tile coding. Your goal is to reach an average reward of -110 during 100 evaluation episodes.

The environment methods are described in the q_learning assignments, with the following changes:

  • The state returned by the env.step method is a list containing weight indices of the current state (i.e., the feature vector of the state consists of zeros and ones, and only the indices of the ones are returned). The action-value function is therefore approximated as a sum of the weights whose indices are returned by env.step.
  • The env.observation_space.nvec returns a list, where the ii-th element is a number of weights used by first ii elements of state. Notably, env.observation_space.nvec[-1] is the total number of weights.

You can start with the q_learning_tiles.py template, which parses several useful parameters and creates the environment. Implementing Q-learning is enough to pass the assignment, even if both N-step Sarsa and Tree Backup converge a little faster. The default number of tiles in tile encoding (i.e., the size of the list with weight indices) is args.tiles=8, but you can use any number you want (but the assignment is solvable with 8).

During evaluation in ReCodEx, three different random seeds will be employed, and you need to reach the required return on all of them. The time limit for each test is 5 minutes.

q_network

 Deadline: Nov 07, 23:59  5 points

Solve the continuous CartPole-v1 environment environment from the OpenAI Gym using Q-learning with neural network as a function approximation.

You can start with the q_network.py template, which provides a simple network implementation in TensorFlow. Feel free to use PyTorch instead, if you like.

The continuous environment is very similar to a discrete one, except that the states are vectors of real-valued observations with shape env.observation_space.shape.

Use Q-learning with neural network as a function approximation, which for a given state returns state-action values for all actions. You can use any network architecture, but one hidden layer of several dozens ReLU units is a good start. Your goal is to reach an average return of 450 during 100 evaluation episodes.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 10 minutes (so you can train in ReCodEx, but you can also pretrain your network if you like).

car_racing

 Deadline: Nov 14, 23:59  8 points + 8 bonus

The goal of this competition is to use Deep Q Networks (and any of Rainbow improvements) on a more real-world CarRacing-v0 environment from the OpenAI Gym.

The supplied car_racing_environment.py provides the environment. It is continuous and states are RGB images of size 96×96×396×96×3, but you can downsample them even more. The actions are also continuous and consist of an array with the following three elements:

  • steer in range [-1, 1]
  • gas in range [0, 1]
  • brake in range [0, 1]; note that full brake is quite aggressive, so you might consider using less force when braking

Internally you should probably generate discrete actions and convert them to the required representation before the step call. The smallest set is probably left, right, gas, brake and no-op, but you can use a more fine-grained one if you like.

The environment also support frame skipping, which improves its performance (only some frames need to be rendered).

In ReCodEx, you are expected to submit an already trained model, which is evaluated on 15 different tracks with a total time limit of 15 minutes. If your average return is at least 300, you obtain 8 points. The task is also a competition and at most 8 points will be awarded according to relative ordering of your solution performances.

The car_racing.py template parses several useful parameters and creates the environment. Note that the car_racing_environment.py can be executed directly and in that case you can drive the car using arrows.

Also, you might want to use a vectorized version of the environment for training, which runs several individual environments in separate processes. The template contains instructions how to create it. The vectorized environment expects a vector of actions and returns a vector of observations, rewards, dones and infos. When one of the environments is done, it is immediately reset and state is the initial state of a new episode.

reinforce

 Deadline: Nov 21, 23:59  4 points

Solve the continuous CartPole-v1 environment environment from the OpenAI Gym using the REINFORCE algorithm.

Your goal is to reach an average return of 490 during 100 evaluation episodes.

Start with the reinforce.py template, which provides a simple network implementation in TensorFlow. Feel free to use PyTorch instead, if you like.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 5 minutes.

reinforce_baseline

 Deadline: Nov 21, 23:59  4 points

This is a continuation of the reinforce assignment.

Using the reinforce_baseline.py template, solve the continuous CartPole-v1 environment environment using the REINFORCE with baseline algorithm.

Using a baseline lowers the variance of the value function gradient estimator, which allows faster training and decreases sensitivity to hyperparameter values. To reflect this effect in ReCodEx, note that the evaluation phase will automatically start after 200 episodes. Using only 200 episodes for training in this setting is probably too little for the REINFORCE algorithm, but suffices for the variant with a baseline. In this assignment, you must train your agent in ReCodEx using the provided environment only.

Your goal is to reach an average return of 490 during 100 evaluation episodes.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 5 minutes.

cart_pole_pixels

 Deadline: Nov 21, 23:59  4 points + 5 bonus

The supplied cart_pole_pixels_environment.py generates a pixel representation of the CartPole environment as an 80×8080×80 image with three channels, with each channel representing one time step (i.e., the current observation and the two previous ones).

During evaluation in ReCodEx, three different random seeds will be employed, each with time limit of 10 minutes, and if you reach an average return at least 300 on all of them, you obtain 4 points. The task is also a competition and at most 5 points will be awarded according to relative ordering of your solution performances.

The cart_pole_pixels.py template parses several parameters and creates the environment. You are again supposed to train the model beforehand and submit only the trained neural network.

paac

 Deadline: Nov 28 Dec 05, 23:59  4 points

Solve the CartPole-v1 environment environment using parallel actor-critic algorithm, employing the vectorized environment described in car_racing assignment.

Your goal is to reach an average return of 450 during 100 evaluation episodes.

Start with the paac.py template, which provides a simple network implementation in TensorFlow. Feel free to use PyTorch instead, if you like.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 10 minutes.

paac_continuous

 Deadline: Nov 28 Dec 05, 23:59  5 points

Solve the MountainCarContinuous-v0 environment environment using parallel actor-critic algorithm with continuous actions. When actions are continuous, env.action_space is the same Box space as env.observation_space, offering:

  • env.action_space.shape, which specifies the shape of actions (you can assume actions are always a 1D vector),
  • env.action_space.low and env.action_space.high, which specify the ranges of the corresponding actions.

Your goal is to reach an average return of 90 during 100 evaluation episodes.

Start with the paac_continuous.py template, which provides a simple network implementation in TensorFlow. Feel free to use PyTorch instead, if you like.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 10 minutes.

ddpg

 Deadline: Nov 28 Dec 05, 23:59  6 points

Solve the Pendulum-v0 environment environment using deep deterministic policy gradient algorithm. The environment is continuous, states and actions are described at OpenAI Gym Wiki.

Your goal is to reach an average return of -200 during 100 evaluation episodes.

Start with the ddpg.py template, which provides a simple network implementation in TensorFlow. Feel free to use PyTorch instead, if you like.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the required return on all of them. Time limit for each test is 10 minutes.

walker

 Deadline: Dec 05 Dec 12, 23:59  5 points

In this exercise exploring continuous robot control, try solving the BipedalWalker-v3 environment environment from the OpenAI Gym. The environment is continuous, states and actions are described at OpenAI Gym Wiki.

Note that the penalty of -100 on crash makes the training considerably slower. Even if all of DDPG, TD3 and SAC can be trained with original rewards, overriding the reward at the end of episode to 0 speeds up training considerably.

In ReCodEx, you are expected to submit an already trained model, which is evaluated with two seeds, each for 100 episodes with a time limit of 10 minutes. If your average return is at least 200, you obtain 5 points.

The walker.py template contains the skeleton for implementing the SAC agent, but you can also solve the assignment with DDPG/TD3.

walker_hardcore

 Deadline: Dec 05 Dec 12, 23:59  6 points + 8 bonus

As an extension of the walker assignment, solve the BipedalWalkerHardcore-v3 environment environment from the OpenAI Gym.

Note that the penalty of -100 on crash can discourage or even stop training, so overriding the reward at the end of episode to 0 (or descresing it substantially) makes the training considerably easier (I have not surpassed return 0 with neither TD3 nor SAC with the original -100 penalty).

In ReCodEx, you are expected to submit an already trained model, which is evaluated with three seeds, each for 100 episodes with a time limit of 10 minutes. If your average return is at least 100, you obtain 6 points. The task is also a competition and at most 8 points will be awarded according to relative ordering of your solution performances.

The walker_hardcore.py template shows a basic structure of evaluaton in ReCodEx, but you most likely want to start either with ddpg.py. or with walker.py and just change the env argument to BipedalWalkerHardcore-v3.

brax_cheetah

 Deadline: Dec 19, 23:59  4 points; not required for passing the exam with grade 1 by solving all assignments

In this optional exercise, try using the DDPG/TD3/SAC algorithm to solve the HalfCheetah environment, but using the Halfcheetah environment from the Brax engine.

You will need additional packages for this assignment, namely brax==0.0.8 jax==0.2.25 jaxlib==0.1.74 typing-extensions~=3.7.4, where the versions are chosen to be compatible with the other course packages. Unfortunately, the binary packages are available only for Linux and OS X; the Windows users should use Windows Subsystem for Linux according to the JAX installation instructions.

The template brax_cheetah.py shows how to

  • create a single Brax environment
  • create a vectorized Brax environment
  • render an episode in the Brax HTML visualizer

In ReCodEx, you are expected to submit an already trained model, which is evaluated with two seeds, each for 100 episodes with a time limit of 10 minutes. If your average return is at least 5000 on all of them, you obtain 4 bonus points.

trace_algorithms

 Deadline: Dec 19, 23:59  4 points

Starting with the trace_algorithms.py template, implement the following state value estimations:

  • use nn-step estimates for a given nn;
  • if requested, use eligibility traces with a given λλ;
  • allow off-policy correction using importance sampling with control variates, optionally clipping the individual importance sampling ratios by a given threshold.

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 trace_algorithms.py --n=1
Episode 100, mean 100-episode return -96.61 +-96.27
Episode 200, mean 100-episode return -30.95 +-57.27
Episode 300, mean 100-episode return -27.00 +-47.28
Episode 400, mean 100-episode return -11.97 +-32.78
Episode 500, mean 100-episode return -10.74 +-34.76
Episode 600, mean 100-episode return -6.70 +-32.46
Episode 700, mean 100-episode return -3.87 +-18.23
Episode 800, mean 100-episode return 0.72 +-11.11
Episode 900, mean 100-episode return -1.34 +-24.02
Episode 1000, mean 100-episode return 2.68 +-9.17
The mean 1000-episode return after evaluation -38.58 +-87.25
  • python3 trace_algorithms.py --n=4
Episode 100, mean 100-episode return -81.76 +-98.03
Episode 200, mean 100-episode return -3.72 +-19.53
Episode 300, mean 100-episode return 0.59 +-10.51
Episode 400, mean 100-episode return 0.99 +-8.76
Episode 500, mean 100-episode return -0.35 +-9.10
Episode 600, mean 100-episode return 1.39 +-8.22
Episode 700, mean 100-episode return 2.42 +-7.80
Episode 800, mean 100-episode return 2.38 +-8.33
Episode 900, mean 100-episode return 2.79 +-7.16
Episode 1000, mean 100-episode return 0.42 +-8.51
The mean 1000-episode return after evaluation -9.03 +-57.13
  • python3 trace_algorithms.py --n=8
Episode 100, mean 100-episode return -107.63 +-113.99
Episode 200, mean 100-episode return -3.57 +-16.96
Episode 300, mean 100-episode return 0.17 +-10.35
Episode 400, mean 100-episode return 0.20 +-8.34
Episode 500, mean 100-episode return 0.27 +-12.30
Episode 600, mean 100-episode return 1.45 +-8.57
Episode 700, mean 100-episode return 2.39 +-8.68
Episode 800, mean 100-episode return 1.92 +-8.32
Episode 900, mean 100-episode return -2.12 +-15.31
Episode 1000, mean 100-episode return -5.06 +-28.00
The mean 1000-episode return after evaluation -69.59 +-100.82
  • python3 trace_algorithms.py --n=4 --trace_lambda=0.6
Episode 100, mean 100-episode return -87.36 +-95.03
Episode 200, mean 100-episode return -10.61 +-28.93
Episode 300, mean 100-episode return -3.48 +-15.93
Episode 400, mean 100-episode return -2.11 +-12.50
Episode 500, mean 100-episode return 1.09 +-8.20
Episode 600, mean 100-episode return 1.40 +-8.85
Episode 700, mean 100-episode return 3.78 +-7.59
Episode 800, mean 100-episode return 1.77 +-8.44
Episode 900, mean 100-episode return 0.53 +-9.03
Episode 1000, mean 100-episode return 1.73 +-7.72
The mean 1000-episode return after evaluation 7.63 +-2.44
  • python3 trace_algorithms.py --n=8 --trace_lambda=0.6
Episode 100, mean 100-episode return -110.14 +-107.12
Episode 200, mean 100-episode return -18.70 +-45.52
Episode 300, mean 100-episode return -4.57 +-23.40
Episode 400, mean 100-episode return 1.17 +-8.73
Episode 500, mean 100-episode return 1.57 +-8.19
Episode 600, mean 100-episode return 2.46 +-8.84
Episode 700, mean 100-episode return 1.47 +-8.32
Episode 800, mean 100-episode return 0.11 +-9.15
Episode 900, mean 100-episode return 1.59 +-8.02
Episode 1000, mean 100-episode return 0.85 +-9.86
The mean 1000-episode return after evaluation 6.63 +-16.25
  • python3 trace_algorithms.py --n=1 --off_policy
Episode 100, mean 100-episode return -74.33 +-71.96
Episode 200, mean 100-episode return -24.48 +-32.66
Episode 300, mean 100-episode return -19.26 +-26.23
Episode 400, mean 100-episode return -10.81 +-22.29
Episode 500, mean 100-episode return -10.40 +-19.60
Episode 600, mean 100-episode return -2.12 +-14.89
Episode 700, mean 100-episode return -3.98 +-17.19
Episode 800, mean 100-episode return -0.89 +-11.64
Episode 900, mean 100-episode return 0.04 +-9.86
Episode 1000, mean 100-episode return 1.02 +-7.64
The mean 1000-episode return after evaluation -22.17 +-73.86
  • python3 trace_algorithms.py --n=4 --off_policy
Episode 100, mean 100-episode return -83.40 +-96.59
Episode 200, mean 100-episode return -19.17 +-59.77
Episode 300, mean 100-episode return 0.92 +-9.35
Episode 400, mean 100-episode return 2.33 +-8.55
Episode 500, mean 100-episode return 0.06 +-19.31
Episode 600, mean 100-episode return 0.54 +-9.05
Episode 700, mean 100-episode return 3.27 +-7.52
Episode 800, mean 100-episode return 3.15 +-7.68
Episode 900, mean 100-episode return 1.27 +-8.25
Episode 1000, mean 100-episode return 1.93 +-7.03
The mean 1000-episode return after evaluation 7.87 +-2.75
  • python3 trace_algorithms.py --n=8 --off_policy
Episode 100, mean 100-episode return -114.45 +-108.68
Episode 200, mean 100-episode return -35.19 +-85.26
Episode 300, mean 100-episode return -16.64 +-65.03
Episode 400, mean 100-episode return -24.48 +-80.72
Episode 500, mean 100-episode return -21.21 +-75.67
Episode 600, mean 100-episode return -0.48 +-16.61
Episode 700, mean 100-episode return 0.77 +-14.33
Episode 800, mean 100-episode return 2.67 +-7.60
Episode 900, mean 100-episode return 3.60 +-6.91
Episode 1000, mean 100-episode return 2.40 +-7.14
The mean 1000-episode return after evaluation 7.24 +-11.67
  • python3 trace_algorithms.py --n=1 --off_policy --vtrace_clip=1
Episode 100, mean 100-episode return -68.42 +-71.79
Episode 200, mean 100-episode return -29.45 +-42.22
Episode 300, mean 100-episode return -18.72 +-26.28
Episode 400, mean 100-episode return -13.66 +-27.16
Episode 500, mean 100-episode return -6.64 +-19.08
Episode 600, mean 100-episode return -3.34 +-14.91
Episode 700, mean 100-episode return -5.72 +-16.65
Episode 800, mean 100-episode return -0.49 +-11.18
Episode 900, mean 100-episode return -0.67 +-10.13
Episode 1000, mean 100-episode return -0.11 +-11.14
The mean 1000-episode return after evaluation -18.13 +-69.41
  • python3 trace_algorithms.py --n=4 --off_policy --vtrace_clip=1
Episode 100, mean 100-episode return -82.58 +-89.97
Episode 200, mean 100-episode return -8.99 +-28.13
Episode 300, mean 100-episode return 1.72 +-9.70
Episode 400, mean 100-episode return 1.52 +-8.17
Episode 500, mean 100-episode return 0.99 +-6.85
Episode 600, mean 100-episode return 1.84 +-8.79
Episode 700, mean 100-episode return 3.36 +-8.10
Episode 800, mean 100-episode return 3.17 +-7.44
Episode 900, mean 100-episode return 2.75 +-7.18
Episode 1000, mean 100-episode return 0.62 +-8.92
The mean 1000-episode return after evaluation 7.85 +-2.62
  • python3 trace_algorithms.py --n=8 --off_policy --vtrace_clip=1
Episode 100, mean 100-episode return -94.76 +-99.63
Episode 200, mean 100-episode return -3.98 +-27.01
Episode 300, mean 100-episode return 1.55 +-9.64
Episode 400, mean 100-episode return 0.69 +-8.90
Episode 500, mean 100-episode return 2.73 +-7.47
Episode 600, mean 100-episode return 3.14 +-7.40
Episode 700, mean 100-episode return 2.25 +-7.79
Episode 800, mean 100-episode return 4.01 +-7.08
Episode 900, mean 100-episode return 2.24 +-8.05
Episode 1000, mean 100-episode return 1.10 +-8.47
The mean 1000-episode return after evaluation 7.93 +-2.62

Note that your results may be slightly different, depending on your CPU type and whether you use a GPU.

  • python3 trace_algorithms.py --episodes=50 --n=1
The mean 1000-episode return after evaluation -189.65 +-45.59
  • python3 trace_algorithms.py --episodes=50 --n=4
The mean 1000-episode return after evaluation -150.78 +-88.82
  • python3 trace_algorithms.py --episodes=50 --n=8
The mean 1000-episode return after evaluation -131.20 +-98.26
  • python3 trace_algorithms.py --episodes=50 --n=4 --trace_lambda=0.6
The mean 1000-episode return after evaluation -155.89 +-85.56
  • python3 trace_algorithms.py --episodes=50 --n=8 --trace_lambda=0.6
The mean 1000-episode return after evaluation -166.31 +-77.19
  • python3 trace_algorithms.py --episodes=50 --n=1 --off_policy
The mean 1000-episode return after evaluation -188.56 +-47.90
  • python3 trace_algorithms.py --episodes=50 --n=4 --off_policy
The mean 1000-episode return after evaluation -168.89 +-74.65
  • python3 trace_algorithms.py --episodes=50 --n=8 --off_policy
The mean 1000-episode return after evaluation -171.46 +-71.94
  • python3 trace_algorithms.py --episodes=50 --n=1 --off_policy --vtrace_clip=1
The mean 1000-episode return after evaluation -193.59 +-36.46
  • python3 trace_algorithms.py --episodes=50 --n=4 --off_policy --vtrace_clip=1
The mean 1000-episode return after evaluation -163.49 +-79.84
  • python3 trace_algorithms.py --episodes=50 --n=8 --off_policy --vtrace_clip=1
The mean 1000-episode return after evaluation -171.98 +-71.53

az_quiz

 Deadline: Jan 02, 23:59 (competition); Feb 13, 23:49 (regular points)  10 points + 10 bonus

In this competition assignment, use Monte Carlo Tree Search to learn an agent for a simplified version of AZ-kvíz. In our version, the agent does not have to answer questions and we assume that all answers are correct.

The game itself is implemented in the az_quiz.py module, using randomized=False constructor argument.

The evaluation in ReCodEx should be implemented by returning an object implementing a method play, which given an AZ-kvíz instance returns the chosen move. The illustration of the interface is in the az_quiz_player_random.py module, which implements a random agent.

Your solution in ReCodEx is automatically evaluated against a very simple heuristic az_quiz_player_simple_heuristic.py, playing 56 games as a starting player and 56 games as a non-starting player. The time limit for the games is 10 minutes and you should see the win rate directly in ReCodEx. If you achieve at least 80%, you will pass the assignment.

The final competition evaluation will be performed after the deadline by a round-robin tournament. In this tournament, we also consider games where the first move is chosen for the first player (FirstChosen label in ReCodEx, --first_chosen option of the evaluator).

The az_quiz_evaluator.py can be used to evaluate any two given implementations and there are two interactive players available, az_quiz_player_interactive_mouse.py and az_quiz_player_interactive_keyboard.py.

The starting template is available in the az_quiz_agent.py module. Additionally, you might also want to read the official pseudocode for AlphaZero, which however contains several errors:

  • Below line 215, the following line should be inserted
    root.visit_count = 1
    
    Otherwise the visit_count is 0, ucb_score will return all zeros for all actions and during the first simulation, the last valid action in the root will always be chosen.
  • On line 237, next action should be sampled according to a distribution of normalized visit counts, not according to a softmax of visit counts.
  • On line 258, the value of a child should be inverted, if the player to play in the current node is the other one than in the child (which is almost always true). If the assume the values are in [1,1][-1, 1] range, the fixed line should be
    value_score = - child.value()
    
  • On line 279, a value is inverted using 1 - value; however, for values in [1,1][-1, 1], it should be inverted as - value.
  • Below line 287, the sampled gamma random variables should be normalized to produce a Dirichlet random sample:
    noise /= np.sum(noise)
    

az_quiz_randomized

 Deadline: Jan 02, 23:59  5 bonus; not required for passing the exam with grade 1 by solving all assignments

Extend the az_quiz assignment to handle the possibility of wrong answers. Therefore, when choosing a field, the agent might answer incorrectly.

To instantiate this randomized game variant, pass randomized=True to the AZQuiz class of az_quiz.py.

The Monte Carlo Tree Search has to be slightly modified to handle stochastic MDP. The information about distribution of possible next states is provided by the AZQuiz.all_moves method, which returns a list of (probability, az_quiz_instance) next states (in our environment, there are always two possible next states).

memory_game

 Deadline: Feb 13, 23:59  3 points; any non-zero number counts as solved for passing the exam with grade 1 by solving all the assignments

In this exercise we explore a partially observable environment. Consider a one-player variant of a memory game (pexeso), where a player repeatedly flip cards. If the player flips two cards with the same symbol in succession, the cards are removed and the player recieves a reward of +2. Otherwise the player recieves a reward of -1. An episode ends when all cards are removed. Note that it is valid to try to flip an already removed card.

Let there be NN cards in the environment, NN being even. There are N+1N+1 actions – the first NN flip the corresponding card, and the last action flips the unused card with the lowest index (or the card NN if all have been used already). The observations consist of a pair of discrete values (card, symbol), where the card is the index of the card flipped, and the symbol is the symbol on the flipped card. The env.states returns a pair (N,N/2)(N, N/2), representing there are NN card indices and N/2N/2 symbol indices.

Every episode can be ended by at most 3N/23N/2 actions, and the required return is therefore greater or equal to zero. Note that there is a limit of at most 2N2N actions per episode. The described environment is provided by the memory_game_environment.py module.

Your goal is to solve the environment, using supervised learning via the provided expert episodes and networks with external memory. The environment implements an env.expert_episode() method, which returns a fresh correct episode as a list of (state, action) pairs (with the last action being None).

ReCodEx evaluates your solution on environments with 8, 12 and 16 cards (utilizing the --cards argument). For each card number, 100 episodes are simulated once you pass evaluating=True to env.reset and your solution gets 1 point if the average return is nonnegative. You can train the agent directly in ReCodEx (the time limit is 15 minutes), or submit a pre-trained one.

A template memory_game.py is available, commenting a possible use of memory augmented networks.

memory_game_rl

 Deadline: Feb 13, 23:59  5 points; any non-zero number counts as solved for passing the exam with grade 1 by solving all the assignments

This is a continuation of the memory_game assignment.

In this task, your goal is to solve the memory game environment using reinforcement learning. That is, you must not use the env.expert_episode method during training. You can start with the memory_game_rl.py template, which extends the memory_game template by generating training episodes suitable for some reinforcement learning algorithm.

ReCodEx evaluates your solution on environments with 4, 6 and 8 cards (utilizing the --cards argument). For each card number, your solution gets 2 points (1 point for 4 cards) if the average return is nonnegative. You can train the agent directly in ReCodEx (the time limit is 15 minutes), or submit a pre-trained one.

ppo

 Deadline: Feb 27, 23:49  3 points; not required for passing the exam with grade 1 by solving all assignments

Implement the PPO algorithm in a single-agent settings. Notably, solve the SingleCollect environment implemented by the multi_collect_environment.py module. To familiarize with it, you can watch a trained agent and you can run the module directly, controlling the agent with the arrow keys. In the environment, your goal is to reach a known place, obtaining rewards based on the agent's distance. If the agent is continuously occupying the place for some period of time, it gets a large reward and the place is moved randomly. The environment runs for 250 steps and it is considered solved if you obtain a return of at least 500.

The ppo.py template contains a skeleton of the PPO algorithm implementation. Regarding the unspecified hyperparameters, I would consider the following ranges:

  • batch_size between 64 and 512
  • clip_epsilon between 0.1 and 0.2
  • epochs between 1 and 10
  • gamma between 0.97 and 1.0
  • trace_lambda is usually 0.95
  • workers between 16 and 128
  • worker_steps between tens and hundreds

My implementation trains in approximately three minutes of CPU time.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the average return of 450 on all of them. Time limit for each test is 10 minutes.

mappo

 Deadline: Feb 27, 23:49  4 points; not required for passing the exam with grade 1 by solving all assignments

Implement MAPPO in a multi-agent settings. Notably, solve a multi-agent extension of SingleCollect with 2 agents, the MultiCollect2 environment implemented again by the multi_collect_environment.py module (you can watch the trained agents). The environment is a generalization of SingleCollect. If there are AA agents, there are also AA target places, and each place rewards the closest agent. Additionally, any agent colliding with others gets a negative reward, and the environment reward is the average of the agents' rewards (to keep the rewards less dependent on the number of agents). Again, the environment runs for 250 steps and is considered solved when reaching return of at least 500.

The mappo.py template contains a skeleton of the MAPPO algorithm implementation. I use hyperparameter values quite similar to the ppo assignment, with a notable exception of a smaller learning_rate=3e-4, which is already specified in the template.

My implementation successfully converges in only circa 50% of the cases, and trains in roughly 10-20 minutes.

During evaluation in ReCodEx, two different random seeds will be employed, and you need to reach the average return of 450 on all of them. Time limit for each test is 10 minutes.

Submitting to ReCodEx

When submitting a competition solution to ReCodEx, you should submit a trained agent and a Python source capable of running it.

Furthermore, please also include the Python source and hyperparameters you used to train the submitted model. But be careful that there still must be exactly one Python source with a line starting with def main(.

Do not forget about the maximum allowed model size and time and memory limits.

Competition Evaluation

  • Before the deadline, ReCodEx prints the exact performance of your agent, but only if it is worse than the baseline.

    If you surpass the baseline, the assignment is marked as solved in ReCodEx and you immediately get regular points for the assignment. However, ReCodEx does not print the reached performance.

  • After the competition deadline, the latest submission of every user surpassing the required baseline participates in a competition. Additional bonus points are then awarded according to the ordering of the performance of the participating submissions.

  • After the competition results announcement, ReCodEx starts to show the exact performance for all the already submitted solutions and also for the solutions submitted later.

What Is Allowed

  • Unless stated otherwise, you can use any algorithm to solve the competition task at hand, but the implementation must be created by you.
  • Both TensorFlow and PyTorch are available in ReCodEx (but there are no GPUs).

Install

  • Installing to central user packages repository

    You can install all required packages to central user packages repository using pip3 install --user tensorflow==2.6.2 tensorflow_probability==0.14.1 numpy==1.19.5 gym==0.20.0 pyglet==1.5.21 ufal.pybox2d==2.3.10.

  • Installing to a virtual environment

    Python supports virtual environments, which are directories containing independent sets of installed packages. You can create a virtual environment by running python3 -m venv VENV_DIR followed by VENV_DIR/bin/pip3 install tensorflow==2.6.2 tensorflow_probability==0.14.1 numpy==1.19.5 gym==0.20.0 pyglet==1.5.21 ufal.pybox2d==2.3.10.

ReCodEx

  • What files can be submitted to ReCodEx?

    You can submit multiple files of any type to ReCodEx. There is a limit of 20 files per submission, with a total size of 20MB.

  • What file does ReCodEx execute and what arguments does it use?

    Exactly one file with py suffix must contain a line starting with def main(. Such a file is imported by ReCodEx and the main method is executed (during the import, __name__ == "__recodex__").

    The file must also export an argument parser called parser. ReCodEx uses its arguments and default values, but it overwrites some of the arguments depending on the test being executed – the template should always indicate which arguments are set by ReCodEx and which are left intact.

  • What are the time and memory limits?

    The memory limit during evaluation is 1.5GB. The time limit varies, but it should be at least 10 seconds and at least twice the running time of my solution.

  • Do agents need to be trained directly in ReCodEx?

    No, you can pre-train your agent locally (unless specified otherwise in the task description).

Requirements

To pass the practicals, you need to obtain at least 80 points, excluding the bonus points. Note that all surplus points (both bonus and non-bonus) will be transfered to the exam. In total, assignments for at least 120 points (not including the bonus points) will be available, and if you solve all the assignments (either before the first or the second deadline), you automatically pass the exam with grade 1.

To pass the exam, you need to obtain at least 60, 75 and 90 out of 100-point exam, to obtain grades 3, 2 and 1, respectively. (PhD students with binary grades require 75 points.) The exam consists of 100-point-worth questions from the list below (the questions are randomly generated, but in such a way that there is at least one question from every lecture). In addition, you can get surplus points from the practicals and at most 10 points for community work (i.e., fixing slides or reporting issues) – but only the points you already have at the time of the exam count. You can take the exam without passing the practicals first.

Exam Questions

Lecture 1 Questions

  • Derive how to incrementally update a running average (how to compute an average of NN numbers using the average of the first N1N-1 numbers). [5]

  • Describe multi-arm bandits and write down the ϵ\epsilon-greedy algorithm for solving it. [5]

  • Define a Markov Decision Process, including the definition of a return. [5]

  • Describe how does a partially observable Markov decision process extend the Markov decision process and how is the agent altered. [5]

Lecture 2 Questions

  • Define a value function, such that all expectations are over simple random variables (actions, states, rewards), not trajectories. [5]

  • Define an action-value function, such that all expectations are over simple random variables (actions, states, rewards), not trajectories. [5]

  • Express a value function using an action-value function, and express an action-value function using a value function. [5]

  • Define optimal value function and optimal action-value function. Then define optimal policy in such a way that its existence is guaranteed. [5]

  • Write down the Bellman optimality equation. [5]

  • Define the Bellman backup operator. [5]

  • Write down the value iteration algorithm. [5]

  • Define the supremum norm ||\cdot||_\infty and prove that Bellman backup operator is a contraction with respect to this norm. [10]

  • Formulate and prove the policy improvement theorem. [10]

  • Write down the policy iteration algorithm. [10]

  • Write down the tabular Monte-Carlo on-policy every-visit ϵ\epsilon-soft algorithm. [10]

Lecture 3 Questions

  • Write down the Sarsa algorithm. [10]

  • Write down the Q-learning algorithm. [10]

  • Write down the Double Q-learning algorithm. [10]

  • Elaborate on how can importance sampling estimate expectations with respect to π\pi based on samples of bb. [5]

  • Show how to estimate returns in the off-policy case, both with (1) ordinary importance sampling and (2) weighted importance sampling. [10]

  • Write down the Expected Sarsa algorithm and show how to obtain Q-learning from it. [10]

  • Show the bootstrapped estimate of nn-step return. [5]

  • Write down the update in on-policy nn-step Sarsa (assuming you already have nn previous steps, actions and rewards). [5]

  • Write down the update in off-policy nn-step Sarsa with importance sampling (assuming you already have nn previous steps, actions and rewards). [10]

  • Write down the update of nn-step Tree-backup algorithm (assuming you already have nn previous steps, actions and rewards). [10]

Lecture 4 Questions

  • Assuming function approximation, define Mean squared value error. [5]

  • Write down the gradient Monte-Carlo on-policy every-visit ϵ\epsilon-soft algorithm. [10]

  • Write down the semi-gradient ϵ\epsilon-greedy Sarsa algorithm. [10]

  • Prove that semi-gradient TD update is not an SGD update of any loss. [10]

  • What are the three elements causing off-policy divergence with function approximation? Write down the Baird's counterexample. [10]

  • Explain the role of a replay buffer in Deep Q Networks. [5]

  • How is the target network used and updated in Deep Q Networks? [5]

  • Explain how is reward clipping used in Deep Q Networks. What other clipping is used? [5]

  • Formulate the loss used in Deep Q Networks. [5]

  • Write down the Deep Q Networks training algorithm. [10]

Lecture 5 Questions

  • Explain the difference between DQN and Double DQN. [5]

  • Describe prioritized replay (how are transitions sampled from the replay buffer, how up-to-date are the priorities [according to which we sample], how are unseen transitions boosted, how is importance sampling used to account for the change in the sampling distribution). [10]

  • How is the action-value function computed in dueling networks? [5]

  • Describe a fully connected layer in Noisy nets (parametrization, computation, effective noise generation). [5]

  • In Distributional RL, describe how is the distribution of rewards represented and how it is predicted using a neural network. [5]

  • Write down the distributional Bellman equation, desctibe how are the atom probabilities of a reward distribution modeled, and write down the loss used to train a distributional Q network (including the mapping of atoms, which does not need to be mathematically flawless -- it is enough to describe how it should be done). [10]

Lecture 6 Questions

  • Formulate the policy gradient theorem. [5]

  • Prove the part of the policy gradient theorem showing the value of θvπ(s)\nabla_{\boldsymbol\theta} v_\pi(s). [10]

  • Assuming the policy gradient theorem, formulate the loss used by the REINFORCE algorithm and show how can its gradient be expressed as an expectation over states and actions. [5]

  • Write down the REINFORCE algorithm. [10]

  • Show that introducing baseline does not influence validity of the policy gradient theorem. [5]

  • Write down the REINFORCE with baseline algorithm. [10]

  • Write down the one-step Actor-critic algorithm. [10]

  • How and why is entropy regularization used in policy gradient algorithms? [5]

  • The Asynchronous advantage actor-critic (A3C) policy may utilize recurrent neural networks. How is the training structured to allow backpropagation through them (would vanilla DQN, vanilla REINFORCE, vanilla actor-critic work with recurrent neural networks)? [5]

  • Explain the difference between a regular Actor-critic and Parallel Advantage Actor Critic algorithms. [5]

Lecture 7 Questions

  • Explain the difference between a regular Actor-critic and Parallel Advantage Actor Critic algorithms. [5]

  • Considering continuous actions modeled by a normal distribution with diagonal covariance, describe how is the policy distribution computed (network architecture, output activation functions) and how does the loss of a simple REINFORCE algorithm look like. [5]

  • Formulate the deterministic policy gradient theorem for θvπ(s)\nabla_{\boldsymbol\theta} v_\pi(s). [5]

  • Formulate the deterministic policy gradient theorem for θJ(θ)\nabla_{\boldsymbol\theta} J(\boldsymbol\theta). [5]

  • Prove the part of the deterministic policy gradient theorem showing the value of θvπ(s)\nabla_{\boldsymbol\theta} v_\pi(s). [10]

  • Write down the critic loss (or its derivative) and the actor policy loss (or its derivative) of the Deep Determinisic Policy Gradients (DDPG) algorithm. Make sure to distinguish the target networks from the ones being trained. [10]

  • How is the return estimated in the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm? [5]

  • Write down the critic loss (or its derivative) and the actor policy loss (or its derivative) of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. Make sure to distinguish the target networks from the ones being trained. [10]

Lecture 8 Questions

  • Write down how is the reward augmented in Soft actor critic, and the definitions of the soft action-value function and the soft (state-)value function. Then, define the modified Bellman backup operator Tπ\mathcal{T}_\pi (be sure to indicate whether you are using the augmented or non-augmented reward), whose repeated application converges to the soft actor-value function qπq_\pi, and prove it. [10]

  • Considering soft policy improvement of a policy π\pi, write down the update formula for the improved policy π\pi', and prove that the soft action-value function of the improved policy is greater or equal to the soft action-value function of the original policy. [10]

  • Write down how are the critics and target critics updated in the Soft actor critic algorithm. [5]

  • Write down how is the actor updated in the Soft actor critic algorithm, including the policy reparametrization trick. [5]

  • Regarding the entropy penalty coefficient α\alpha in the Soft actor critic, define what contrained optimization problem we are solving, what is the corresponding Lagrangian (and whether we are minimizing/maximizing it with respect to the policy and α\alpha), and what does the α\alpha update looks like. [5]

  • Define a one-step TD error and express the nn-step return as a sum of them. [5]

  • Define a one-step TD error and express the nn-step return with off-policy correction using control variates as a sum of TD errors. [5]

Lecture 9 Questions

  • Define the λ\lambda-return. [5]

  • Define the nn-step truncated λ\lambda-return. [5]

  • Define a one-step TD error and express the nn-step truncated λ\lambda-return as a sum of them. [5]

  • Define a one-step TD error and express the nn-step truncated λ\lambda-return with off-policy correction as a sum of them. [5]

  • Define the V-trace estimate and write down the policy to whose value function the V-trace estimate converges to. [10]

  • Explain why the fixed point of the V-trace operator does not depend on the truncation of all but the last importance sampling ratios. [10]

  • Write down the critic loss (or its derivative) and the actor policy loss (or its derivative) of the IMPALA algorithm, including the V-trace formula. [10]

  • Sketch the population based training used in the IMPALA algorithm. [5]

  • In PopArt normalization, the value function is computed based on a normalized value predictor nn as σn+μ\sigma n + \mu. Describe how to maintain σ\sigma and μ\mu, how to compute normalized advantage based on return GG, and how is the normalized value predictor modified when the estimates of σ\sigma and μ\mu change. [10]

  • Define the transformed Bellman operator. [5]

  • Define the transformed Bellman operator. Then, assuming hh is strictly monotonically increasing function and considering a deterministic Markov decision process, show to what does a transformed Bellman operator Th\mathcal{T}_h converge and prove it. [10]

  • Write down the return transformation used for Atari environments (for example by R2D2). [5]

  • Describe the replay buffer elements in R2D2. What is the difference between the zero-state and stored-state strategies, and how is burn-in used? [5]

  • Write down the Retrace operator and describe the three possibilities of setting the traces ctc_t: importance sampling, Tree-backup(λ\lambda) and Retrace(λ\lambda). [10]

Lecture 10 Questions

  • Considering multi-arm bandits, write down the UCB algorithm. [5]

  • Describe the inputs and outputs of a neural network used in AlphaZero, and describe the inputs and outputs of a Monte-Carlo tree search. [5]

  • Write down the loss used in AlphaZero algorithm. [5]

  • What quantities are kept in a node of a Monte-Carlo tree search? [5]

  • How are actions selected in a Monte-Carlo tree search? [10]

  • What does AlphaZero use to maintain exploration in a Monte-Carlo tree search? [5]

  • Describe the backup phase of Monte-Carlo tree search, i.e., the steps you perform when you reach a leaf during the tree search. [5]

  • How are the actions selected in AlphaZero self-play? [5]

Lecture 11 Questions

  • Describe the three components of a MuZero model, and describe/draw how they are used to traverse the MCTS tree. [5]

  • Describe the MCTS in MuZero – action selection (including the exact action-values used), how are the three components of a MuZero model used during the tree traversal and leaf evaluation, and the updates during the backup phase. [10]

  • Assuming we already have a filled replay buffer, describe the MuZero training – the losses and the target values used in them. [10]

  • Describe the components of a typical latent-space model in PlaNet (the transition, observation and reward functions, the encoder) and the components of a recurrent state-space model (RSSM). [5]

  • Derive the variational lower bound on logp(o1:Ta1:T)\log p(o_{1:T} | a_{1:T}) used in PlaNet (you can utilize the Jensen's inequality logE[x]E[logx]\log \mathbb{E} [x] \ge \mathbb{E} [\log x]). [10]

Lecture 12 Questions

  • Consider a model with a discrete categorical latent variable z\boldsymbol z sampled from p(z;θ)p(\boldsymbol z; \boldsymbol \theta), with a loss L(z;ω)L(\boldsymbol z; \boldsymbol \omega). Describe how we compute the derivative of the loss LL with respect to the parameters θ\boldsymbol \theta using (a) a straight-through estimator, and (b) a REINFORCE-like gradient estimator with a baseline. [5]

  • Describe the Gumbel-Max trick; in other words, write down how to perform sampling from a categorical distribution using an argmax\operatorname{argmax}, including the procedure for sampling from the Gumbel(0,1)\operatorname{Gumbel}(0, 1) distribution. [5]

  • Consider a discrete categorical variable sampled from logits l\boldsymbol l. Define the Gumbel-softmax(l,T)\operatorname{Gumbel-softmax}(\boldsymbol l, T) distribution with logits l\boldsymbol l and a temperature TT (no need to describe sampling from Gumbel(0,1)\operatorname{Gumbel}(0, 1)), and describe the main difference between the Gumbel-softmax(l,T)\operatorname{Gumbel-softmax}(\boldsymbol l, T) and the softmax(l)\operatorname{softmax}(\boldsymbol l) distributions. [5]

  • Consider a model with a discrete categorical latent variable z\boldsymbol z sampled from p(z;θ)p(\boldsymbol z; \boldsymbol \theta), with a loss L(z;ω)L(\boldsymbol z; \boldsymbol \omega). Describe how we compute the derivative of the loss LL with respect to the parameters θ\boldsymbol \theta using (a) a Gumbel-softmax estimator, and (b) a straight-through Gumbel-softmax estimator. [5]

  • Write down an algorithm implementing a straight-through estimator using automatic differentiation (i.e., in TensorFlow or Pytorch). [5]

  • Describe the six components of the DreamerV2 recurrent state-space model (RSSM). [5]

  • Explain the KL balancing used in DreamerV2. [5]

  • Describe the training of both a critic and an actor in DreamerV2 (including the explicit losses). [10]