How Resilient are Imitation Learning Methods to Sub-Optimal Experts?
Trajectories used in How Resilient are Imitation Learning Methods to Sub-Optimal Experts?
These trajectories are formed by using Stable Baselines. Each file is a dictionary of a set of trajectories with the following keys:
- actions: the action in the given timestamp
t - obs: current state in the given timestamp
t - rewards: reward retrieved after the action in the given timestamp
t - episode_returns: The aggregated reward of each episode (each file consists of 5000 runs)
- episode_Starts: Whether that
obsis the first state of an episode (boolean list).
The code that uses this data is on GitHub: https://github.com/NathanGavenski/How-resilient-IL-methods-are