Active Inference Capsule

Repository for the paper Online reinforcement learning with sparse rewards through an active inference capsule, submitted to NeurIPS2021.

Requirements

Python 3.5+
Pytorch (tested on 1.4.0)
numpy
gym
tqdm
matplotlib
ffmpeg (only if generating videos)

Training a single agent

python run_capsule_mountain_car.py --settings='./paper_results/settings_learned_prior_H5.json' --save_dirpath='./paper_results/simulation_results/'

This will train a single agent and save the model in ./paper_results/simulation_results/

By default, the training progress will be printed. There are setting files for other agents in ./paper_results

Note: the folder ./paper_results/simulation_results/ already contains the results used in the paper.

Training a batch of agents

python run_capsule_mountain_car.py --settings='./paper_results/settings_learned_prior_H5.json' --save_dirpath='./paper_results/simulation_results/' --batch_agents=30

This will train 30 agents and save the models and training statistics in ./paper_results/simulation_results/

By default, this will use all available CPU cores. It can be changed by the option --max_cpu.

The results can be visualized using the script ./mountain_car/plot_multiple_trainings.py or by re-making the figures with ./paper_results/make_figures.py (this will also generate other figures than the training plots).

Generating a video

First, train a single agent. With the option --save_all_episodes=True when training the single agent, the model is saved per episode instead of overriding the latest one. This facilitates making videos of different episodes of the same agent.

python run_capsule_mountain_car.py --settings='./paper_results/settings_learned_prior_H5.json' --save_dirpath='./paper_results/simulation_results/' --make_video=True --model_load_filepath='./paper_results/simulation_results/model_learned_prior_H5.pt'

This will start saving individual frames in a subdirectory of save_dirpath. When the agent reaches the goal, it will call ffmpeg to generate a .mp4 file.

This process is much slower than training because matplotlib takes quite some time to render each frame.

Other