vluzko / causal_rl Goto Github PK

Functions that produce interesting non-stable causal relationships. They need to be relatively easy to learn for neural networks, not produce the same causal relationships repeatedly, and not produce extremely sparse causal graphs.

The causal graph isn't guaranteed to be unique to a state, so it can't always be computed afterwards.

Use to cluster states based on which model works well for them. This works as an alternative/addition to using a buffer to decide what states are "bad" for a particular model.

It's not obvious how to implement this, but there are a few additional constraints we might want on the graph predictor:

• "Similar" states producing similar graphs
• Graphs not changing much between steps

Better to rewrite the environment in mujoco.

Looks like one extra time step is included