Hands-On Reinforcement Learning for Games

Behavioral cloning

Behavioral cloning is sometimes also referred to as imitation learning. While not exactly both the same, we will use the terms interchangeably here. In RL, we use the term sparse rewards or rewards sparsity for any environment where it is difficult for an agent to just finish a task by trial and error and perhaps luck. The larger an environment is, the more sparse the rewards and in many cases, the observation space can be so large that any hope of training an agent at all is extremely difficult. Fortunately, a method called behavioral cloning or imitation learning can solve the problem of sparse rewards by using the observations of humans as previous sampled observations. Unity provides three methods to generate previous ...