pai - Model-based Approximate Reinforcement Learning

model-based reinforcement learning

We face three main challenges in model-based reinforcement learning. First, given a fixed model, we need to perform planning to decide on which actions to play. Second, we need to learn models f and r accurately and efficiently. Third, we need to effectively trade exploration and exploitation.

Planning

Deterministic Dynamics

To begin with, let us assume that our dynamics model is deterministic and known. That is, given a state-action pair, we know the subsequent state.