Hands-On Reinforcement Learning for Games

Training generalized reinforcement learning agents

We often have to remind ourselves that RL is just a derivation of data science best practices and we often have to consider how could you fix a training issue with data science. In the case of RL, we see the same issues we see in data science and machine learning only at different scales and exposed in another manner. One example of this is when an agent is overfitted to an environment that we then try to apply to other general variations of that environment. For instance, imagine the Frozen Lake environment that could be various sizes or even provide random starting points or other variations. By introducing these types of variations, we allow our agent to better generalize to a wider variety ...