Automated synthesis of action selection policies for unmanned vehicles operating in adverse environments

We address the problem of automated action selection policy synthesis for unmanned vehicles operating in adverse environments. We introduce a new evolutionary computation-based approach using which an initial version of the policy is automatically generated and then gradually refined by detecting and fixing its shortcomings. The synthesis technique consists of the automated extraction of the vehicle’s exception states and Genetic Programming (GP) for automated composition and optimization of corrective sequences of commands in the form of macro-actions to be applied locally. The focus is specifically on automated synthesis of a policy for Unmanned Surface Vehicle (USV) to efficiently block the advancement of an intruder boat toward a valuable target. This task requires the USV to utilize reactive planning complemented by short-term forward planning to generate specific maneuvers for blocking. The intruder is human-competitive and exhibits a deceptive behavior so that the USV cannot exploit regularity in its attacking behavior. We compared the performance of a hand-coded blocking policy to the performance of a policy that was automatically synthesized. Our results show that the performance of the automatically generated policy exceeds the performance of the hand-coded policy and thus demonstrates the feasibility of the proposed approach.