Foundations of Deep Reinforcement Learning: Theory and Practice in Python

17. Transition Function

Now that we have looked at states, actions, and rewards, the last component necessary to make a functioning RL environment is a transition function, also known as a model.

The model of an environment can be programmed or learned. Programmable rules are common, and they can yield environments with various levels of complexity. Chess is perfectly described by a simple set of rules. Robot simulations approximate a robot’s dynamics and its surroundings. Modern computer games can be very complex, but they are still built using programmed game engines.

However, when modeling problems that cannot be efficiently programmed, a model of an environment can be learned instead. For example, contact dynamics in robotics is difficult ...