rl-agents

Implementations of Reinforcement Learning agents based on The Bible of Reinforcement Learning [1]

Multi-armed Bandits

Multi-armed Bandits are implemented with stationary and non-stationary environments using following action-selection methods:

• Static
• Random
• Greedy
• ε-greedy
• Split
• Linear decay ε-greedy
• Optimistic
• UCB
• Gradient (w/ and w/o baseline)

Dynamic Programming

The Dynamic Programming implementation consists of an algebraic solution as well as an random agent with seperate case and in-place iterative solutions.

Monte Carlo Methods

Following agents are already implemented:

• First-Visit / Every-Visit
• On-Policy / Off-Policy

Temporal-Difference Learning

Following Agents are already implemented:

• Sarsa
• Expected Sarsa
• Q-Learning
• Double Q-Learning

References

[1] R. S. Sutton and A. G. Barto, Reinforcement learning: an introduction, Second edition. Cambridge, Massachusetts: The MIT Press, 2018.