Programmable-reinforcement-learning is a project mainly written in Common Lisp, it's free.
Reinforcement learning algorithms constrained by a partial program
This repository includes the ALisp and Concurrent ALisp languages. The basic idea is that we want to make reinforcement learning algorithms more efficient. We do this by writing a "partial program", which is like a regular program with some choices left unspecified, and then apply one of a set of learning algorithms to find the optimal "completion" of that program. The framework is based on research carried out in the research group of Stuart Russell at UC Berkeley, by Ron Parr, David Andre, and Bhaskara Marthi.
See doc/index.html for documentation.