Formal verification of Fuzzy-based XAI for Strategic Combat Game

Abstract

Explainable AI is a topic at the forefront of the field currently for reasons involving human trust in AI, correctness, auditing, knowledge transfer, and regulation. AI that is developed with reinforcement learning (RL) is especially of interest due to the non-transparency of what was learned from the environment. RL AI systems have been shown to be "brittle" with respect to the conditions it can safely operate in, and therefore ways to show correctness regardless of input values are of key interest. One way to show correctness is to verify the system using Formal Methods, known as Formal Verification. These methods are valuable, but costly and difficult to implement, leading most to instead favor other methodologies for verification that may be less rigorous, but more easily implemented. In this work, we show methods for development of an RL AI system for aspects of the strategic combat game Starcraft 2 that is performant, explainable, and formally verifiable. The resulting system performs very well on example scenarios while retaining explainability of its actions to a human operator or designer. In addition, it is shown to adhere to formal safety specifications about its behavior.