/la-mbda

LAMBDA is a model-based reinforcement learning agent that uses Bayesian world models for safe policy optimization

Primary LanguagePythonMIT LicenseMIT

LAMBDA

Idea

By taking a Bayesian perspective, LAMBDA learns a world model and uses it to generate sequences using different posterior samples of the world model parameters. Following that, it chooses and learns from the optimistic sequences how to solve the task and from the pessimistic sequences how to adhere to safety restrictions.

Running and plotting

Install dependencies (this may take more than an hour):

conda create -n lambda python=3.6
conda activate lambda
pip3 install .

Run experiments:

python3 experiments/train.py --log_dir results/point_goal2/314 --environment sgym_Safexp-PointGoal2-v0 --total_training_steps 1000000 --safety

Plot:

python3 experiments/plot.py --data_path results/

where the script expects the following directory tree structure:

results
├── algo1
│   └── environment1
│       └── experiment1
│       └── ...
│   └── ...
└── algo2
    └── environment1
        ├── experiment1
        └── experiment2

Acknowledgement

Dreamer codebase which served as a starting point for this github repo