A model-based approach plans ahead using a representation of how an environment works. In reinforcement learning, that model combines transition probabilities with a reward function; it may be known or learned. section 4.0
For business problems, a model can be built from carefully chosen assumptions, parameters, and expert knowledge. Data-driven methods instead learn patterns from historical data, and combining the two approaches can be practical. section 1.2
The transition function determines how the method plans:
- When the transition function is known, dynamic programming offers a direct way to solve the decision problem. section 4.0
- Methods that bypass the transition function learn from sampled state, action, reward, and next-state experiences; this can require more interaction, computation, and time. section 4.0
