PyReason-Gym: Symbolic logic simulator for Reinforcement Learning

Kaustuv Mukherji, Devendra R. Parkar, Lahari Pokala, Dyuman Aditya, Paulo Shakarian, Clark Dorman
Jul 2, 2023
PDF Code

PyReason-Gym is a new symbolic simulator introduced to enable symbolic reinforcement learning. Why is such a simulator necessary? Well, this let’s us learn interpretable agent policies defacto and enables transfering of policies to other simulation environments. We explored these possibilities and worked on the project with following aims:

  1. Introduce and verify the new PyReason-Gym simulator for symbolic reinforcement learning.
  2. Extend exisiting deep reinforcement learning algorithms to take advantage of symbolic representations provided by PyReason-Gym
  3. Demonstrate ease of incorporating human-level symbolic logic to enable reinforcement learning in non-markovian dynamics(temporally non-markovian), which existing algorithms struggle with
  4. Demonstrate ease of interpreting and transfering learned policies to other simulation environments

More about the simulator and underlying inference engine can be found below:

Reinforcement Learning Symbolic Logic Simulator Transfer Learning
Devendra R. Parkar
Devendra R. Parkar
Research Assistant

My broad research interests include complex systems, human cognition modeling and learning theory

Publications

Scalable Semantic Non-Markovian Simulation Proxy for Reinforcement Learning [In Press]
A temporal and annotated-logic based simulation proxy that enables interpretable reinforcement learning
[Accepted at International Conference on Semantic Computing]
Kaustuv Mukherji, Devendra R. Parkar, Lahari Pokala, Dyuman Aditya, Paulo Shakarian, Clark Dorman
Project Video DOI
Scalable Semantic Non-Markovian Simulation Proxy for Reinforcement Learning [In Press]