Assured Reinforcement Learning for Safety-Critical Applications

Principal Investigator: Dr Radu Calinescu  Co-Investigator: Dr Daniel Kudenko	Type of project: Dstl National UK PhD programme
Project partner:     Defence Science & Technology Laboratory	Start date: October 2013  End date: September 2017

Project summary

Reinforcement learning (RL) has long been used to develop software agents capable of adaptive behaviour in scenarios characterised by incomplete knowledge and non-deterministic change. Successful commercial applications bear witness to the effectiveness of the approach. Nevertheless, this success has so far eluded the important class of safety-critical applications. Despite significant demand and promising preliminary results, RL adoption in safety-critical applications has rarely progressed beyond simulation and testing. Two major limitations of existing RL approaches led to this lack of adoption. First, they are unable to guarantee compliance with requirements, which is a key demand for safety-critical applications. Second, they produce autonomous agents whose actions are difficult to understand by their human stakeholders, and therefore difficult to trust even when correct. Our project will address these limitations, and therefore enable the exploitation of reinforcement learning in the safety-critical domain.

People

PhD student: George Mason

Publications

R. Calinescu (2013) — Emerging Techniques for the Engineering of Self-Adaptive High-Integrity Software. In: Javier Camara, Rogerio de Lemos, Carlo Ghezzi and Antonia Lopes (editors), Assurances for Self-Adaptive Systems, Volume 7740 of Lecture Notes in Computer Science, Springer, pages 297-310.