Programmatically Interpretable Reinforcement Learning

Abhinav Verma, Vijayaraghavan Murali, Rishabh Singh, Pushmeet Kohli, Swarat Chaudhuri
Proceedings of the 35th International Conference on Machine Learning, PMLR 80:5045-5054, 2018.

Abstract

We present a reinforcement learning framework, called Programmatically Interpretable Reinforcement Learning (PIRL), that is designed to generate interpretable and verifiable agent policies. Unlike the popular Deep Reinforcement Learning (DRL) paradigm, which represents policies by neural networks, PIRL represents policies using a high-level, domain-specific programming language. Such programmatic policies have the benefits of being more easily interpreted than neural networks, and being amenable to verification by symbolic methods. We propose a new method, called Neurally Directed Program Search (NDPS), for solving the challenging nonsmooth optimization problem of finding a programmatic policy with maximal reward. NDPS works by first learning a neural policy network using DRL, and then performing a local search over programmatic policies that seeks to minimize a distance from this neural “oracle”. We evaluate NDPS on the task of learning to drive a simulated car in the TORCS car-racing environment. We demonstrate that NDPS is able to discover human-readable policies that pass some significant performance bars. We also show that PIRL policies can have smoother trajectories, and can be more easily transferred to environments not encountered during training, than corresponding policies discovered by DRL.

Cite this Paper

BibTeX
@InProceedings{pmlr-v80-verma18a, title = {Programmatically Interpretable Reinforcement Learning}, author = {Verma, Abhinav and Murali, Vijayaraghavan and Singh, Rishabh and Kohli, Pushmeet and Chaudhuri, Swarat}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {5045--5054}, year = {2018}, editor = {Dy, Jennifer and Krause, Andreas}, volume = {80}, series = {Proceedings of Machine Learning Research}, month = {10--15 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v80/verma18a/verma18a.pdf}, url = {https://proceedings.mlr.press/v80/verma18a.html}, abstract = {We present a reinforcement learning framework, called Programmatically Interpretable Reinforcement Learning (PIRL), that is designed to generate interpretable and verifiable agent policies. Unlike the popular Deep Reinforcement Learning (DRL) paradigm, which represents policies by neural networks, PIRL represents policies using a high-level, domain-specific programming language. Such programmatic policies have the benefits of being more easily interpreted than neural networks, and being amenable to verification by symbolic methods. We propose a new method, called Neurally Directed Program Search (NDPS), for solving the challenging nonsmooth optimization problem of finding a programmatic policy with maximal reward. NDPS works by first learning a neural policy network using DRL, and then performing a local search over programmatic policies that seeks to minimize a distance from this neural “oracle”. We evaluate NDPS on the task of learning to drive a simulated car in the TORCS car-racing environment. We demonstrate that NDPS is able to discover human-readable policies that pass some significant performance bars. We also show that PIRL policies can have smoother trajectories, and can be more easily transferred to environments not encountered during training, than corresponding policies discovered by DRL.} }
Endnote
%0 Conference Paper %T Programmatically Interpretable Reinforcement Learning %A Abhinav Verma %A Vijayaraghavan Murali %A Rishabh Singh %A Pushmeet Kohli %A Swarat Chaudhuri %B Proceedings of the 35th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2018 %E Jennifer Dy %E Andreas Krause %F pmlr-v80-verma18a %I PMLR %P 5045--5054 %U https://proceedings.mlr.press/v80/verma18a.html %V 80 %X We present a reinforcement learning framework, called Programmatically Interpretable Reinforcement Learning (PIRL), that is designed to generate interpretable and verifiable agent policies. Unlike the popular Deep Reinforcement Learning (DRL) paradigm, which represents policies by neural networks, PIRL represents policies using a high-level, domain-specific programming language. Such programmatic policies have the benefits of being more easily interpreted than neural networks, and being amenable to verification by symbolic methods. We propose a new method, called Neurally Directed Program Search (NDPS), for solving the challenging nonsmooth optimization problem of finding a programmatic policy with maximal reward. NDPS works by first learning a neural policy network using DRL, and then performing a local search over programmatic policies that seeks to minimize a distance from this neural “oracle”. We evaluate NDPS on the task of learning to drive a simulated car in the TORCS car-racing environment. We demonstrate that NDPS is able to discover human-readable policies that pass some significant performance bars. We also show that PIRL policies can have smoother trajectories, and can be more easily transferred to environments not encountered during training, than corresponding policies discovered by DRL.
APA
Verma, A., Murali, V., Singh, R., Kohli, P. & Chaudhuri, S.. (2018). Programmatically Interpretable Reinforcement Learning. Proceedings of the 35th International Conference on Machine Learning, in Proceedings of Machine Learning Research 80:5045-5054 Available from https://proceedings.mlr.press/v80/verma18a.html.

Related Material