Supervised imitation learning
In the imitation learning setting, our goal is to mimic the expert. Say, we want to train our agent to drive a car. Instead of training the agent from scratch by having them interact with the environment, we can train them with expert demonstrations. Okay, what are expert demonstrations? An expert demonstrations are a set of trajectories consisting of state-action pairs where each action is performed by the expert.
We can train an agent to mimic the actions performed by the expert in various respective states. Thus, we can view expert demonstrations as training data used to train our agent. The fundamental idea of imitation learning is to imitate (learn) the behavior of an expert.
One of the simplest and most naive ways to perform imitation learning is to treat the imitation learning task as a supervised learning task. First, we collect a set of expert demonstrations, and then we train a classifier to perform the same action performed by the...