WAE-MDPs

Wasserstein Auto-encoded Markov Decision Processes

Official implementation of the Wasserstein Auto-encoded MDP (WAE-MDP) framework. The latter enables the distilation of (any kind of) reinforcement learning policies into simpler controllers which are paired with a discrete, tractable model of the environment (a so called latent space model). The two are provided with bisimulation guarantees, which allow formally verifying the behaviors of the agent operating under the simplified policy. The source code provided allows for replicating the experiments of the paper Wasserstein Auto-encoded MDPs: Formal Verification of Efficiently Distilled RL Policies with Many-sided Guarantees.

Florent Delgrange
Florent Delgrange
Doctoral Researcher in Artificial Intelligence

Related