- inverse reinforcement learning for control through evolutionary strategies
For generating demonstration for toy control problems through trajectory optimization, we use pydrake. This custom colab notebook comprises of the i)setup ii) trajectoy generation iii) Visualisation and iv) other boiler plates. At present it is ready for pendulumn, and one can easily download a zipped folder of the trajectory for a given set of dynamics and simulation config.
This repo is a working implementation of guided cost learning (https://arxiv.org/abs/1603.00448) for discrete action space. Currently, it supports the CartPole-v0 environment.
$ python3.8 main.py
- generating_expert.py: Generates an expert on CartPole, by training vanilla policy gradient, and finally stores trained trajecteries as expert samples at expert_samples.
- experts/PG.py: Implementation of vanilla policy gradient. This is reused at several places.
- main.py: Contains the main implementation of GCL.
