causal_rl

Overview

Overall goal: train MBRL agents that have causal models of environments. The code is split into three parts:

• The environments (environments/). The key point is that these environments are annotated with their causal graphs, making it possible to assess performance at all
• The state models (graph_predictor.py, state_predictor.py, model.py). These predict the current causal graph (from state) and the next state (from current state and graph). models is a wrapper around both.
• MBRL algorithms. Totally unimplemented, because no environment is both causal and RL yet. (Note: probably CausalWorld solves this).

exp.py contains leftover code from my thesis. It's in the process of being refactored, but for now it contains examples of running the full code.

Data goes in data/. Logs go in logs/. Plots go in plots/. Most likely this will change and logs/plots will get handled by wandb or a similar service.

Architecture

• Environments: Produce a state, and a causal graph associated with that state. The causal graph is not returned to the learner but does get used for scoring after the fact (i.e. it's not part of the training loop)
• Models: Are composed of two parts. The first (the graph predictor) takes state as input and produces a prediction of the causal graph. The second (the state predictor) takes state and the predicted causal graph as input, and outputs the next state.
  • Graph predictors are maps from $reals^k -> reals^(k^2)$. Entry $i, j$ of the result matrix is the probability that $state[i]$ at the current time step has a causal influence on $state[j]$ of the next time step
  • State predictors are maps from $reals^k x reals^(k^2) -> reals^k$.

Usage

exp.py contains scripts for running several different experiments. Output is stored in plots and runs.

Installation

All tests are being performed with Python 3.8.5. You will need at least Python 3.6 to handle the type annotations.

You will need all the dependencies for PyTorch, particularly a CUDA-enabled GPU. The code may work on a CPU, but it will be very slow and it's untested.

• Run pip install -r requirements.txt
• Run pip install .