Python replication for Sutton & Barto's book Reinforcement Learning: An Introduction (2nd Edition)

If you have any confusion about the code or want to report a bug, please open an issue instead of emailing me directly. And unfortunately I do not have exercise answers for the book.

1. Tic-Tac-Toe

1. Figure 2.1: An exemplary bandit problem from the 10-armed testbed
2. Figure 2.2: Average performance of epsilon-greedy action-value methods on the 10-armed testbed
3. Figure 2.3: Optimistic initial action-value estimates
4. Figure 2.4: Average performance of UCB action selection on the 10-armed testbed
5. Figure 2.5: Average performance of the gradient bandit algorithm
6. Figure 2.6: A parameter study of the various bandit algorithms

1. Figure 3.2: Grid example with random policy
2. Figure 3.5: Optimal solutions to the gridworld example

1. Figure 4.1: Convergence of iterative policy evaluation on a small gridworld
2. Figure 4.2: Jack’s car rental problem
3. Figure 4.3: The solution to the gambler’s problem

1. Figure 5.1: Approximate state-value functions for the blackjack policy
2. Figure 5.2: The optimal policy and state-value function for blackjack found by Monte Carlo ES
3. Figure 5.3: Weighted importance sampling
4. Figure 5.4: Ordinary importance sampling with surprisingly unstable estimates

1. Example 6.2: Random walk
2. Figure 6.2: Batch updating
3. Figure 6.3: Sarsa applied to windy grid world
4. Figure 6.4: The cliff-walking task
5. Figure 6.6: Interim and asymptotic performance of TD control methods
6. Figure 6.7: Comparison of Q-learning and Double Q-learning

1. Figure 7.2: Performance of n-step TD methods on 19-state random walk

1. Figure 8.2: Average learning curves for Dyna-Q agents varying in their number of planning steps
2. Figure 8.4: Average performance of Dyna agents on a blocking task
3. Figure 8.5: Average performance of Dyna agents on a shortcut task
4. Example 8.4: Prioritized sweeping significantly shortens learning time on the Dyna maze task
5. Figure 8.7: Comparison of efficiency of expected and sample updates
6. Figure 8.8: Relative efficiency of different update distributions

1. Figure 9.1: Gradient Monte Carlo algorithm on the 1000-state random walk task
2. Figure 9.2: Semi-gradient n-steps TD algorithm on the 1000-state random walk task
3. Figure 9.5: Fourier basis vs polynomials on the 1000-state random walk task
4. Figure 9.8: Example of feature width’s effect on initial generalization and asymptotic accuracy
5. Figure 9.10: Single tiling and multiple tilings on the 1000-state random walk task

1. Figure 10.1: The cost-to-go function for Mountain Car task in one run
2. Figure 10.2: Learning curves for semi-gradient Sarsa on Mountain Car task
3. Figure 10.3: One-step vs multi-step performance of semi-gradient Sarsa on the Mountain Car task
4. Figure 10.4: Effect of the alpha and n on early performance of n-step semi-gradient Sarsa
5. Figure 10.5: Differential semi-gradient Sarsa on the access-control queuing task

1. Figure 11.2: Baird's Counterexample
2. Figure 11.6: The behavior of the TDC algorithm on Baird’s counterexample
3. Figure 11.7: The behavior of the ETD algorithm in expectation on Baird’s counterexample

1. Figure 12.3: Off-line λ-return algorithm on 19-state random walk
2. Figure 12.6: TD(λ) algorithm on 19-state random walk
3. Figure 12.8: True online TD(λ) algorithm on 19-state random walk
4. Figure 12.10: Sarsa(λ) with replacing traces on Mountain Car
5. Figure 12.11: Summary comparison of Sarsa(λ) algorithms on Mountain Car

1. Example 13.1: Short corridor with switched actions
2. Figure 13.1: REINFORCE on the short-corridor grid world
3. Figure 13.2: REINFORCE with baseline on the short-corridor grid-world

• python 3.6
• numpy
• matplotlib
• seaborn
• tqdm

All files are self-contained

python any_file_you_want.py

If you want to contribute some missing examples or fix some bugs, feel free to open an issue or make a pull request.

Following are missing figures/examples:

• Figure 12.14: The effect of λ