How Imitation Learning?

Generally, Imitation learning is useful when it is easier for an expert to demonstrate the desired behavior so that we don’t have to specify any reward function.

Let’s say we collect expert demonstrations (also known as trajectories in RL)

τ = (s0, a0, s1, a1……) where actions (As) are based on expert’s policy(say human brain). In some cases, we may require “expert” during training.

Once we get this trajectory, we slice this time steps to get pairs of Ss and As.

We then treat these pairs as i.i.d examples and apply Supervised learning.

Changing loss function and optimization strategies in this learning define various imitation learning algorithms. Let’s look at basic ones in them.

#deep-learning #imitation-learning #reinforcement-learning #machine-learning