Tabular Agents

Tabular agents implement reinforcement learning algorithms that store and update action-value estimates in explicit tables. These methods are well-suited for small, discrete state and action spaces, where values can be represented directly without the need for function approximation. They provide a clear and educational foundation for understanding how agents learn from experience.

RLForge includes the following tabular agents:

• SARSA Agent — learns action values using the on-policy SARSA algorithm, updating estimates based on the action actually taken in the next state.

• Q Agent — implements the classic Q-learning algorithm, an off-policy method that updates values toward the greedy action in the next state.

• Expected SARSA Agent — a variant of SARSA that uses the expected value of all possible actions in the next state, weighted by the policy’s probabilities, leading to smoother updates.

These agents are ideal for experimenting with fundamental reinforcement learning concepts and serve as building blocks for more advanced methods.

• Planning Agent
• SARSA Agent
• Q Learning Agent
• Expected SARSA Agent