Chaos engineering: stress-testing algorithms to facilitate resilient strategic military planning.

We present our Chaos Engineering framework that accelerates the emergence of resilient strategic plans to maintain high levels of performance given the turbulent and unpredictable conditions in real-world military deployments. Our proposed Chaos Engine will stress agents by intelligently searching over scenario chaos factors reflecting real-world events such as weather changes, loss of communication, basing and access, platform failures, and enemy force composition and mission plans. Our work is split into four technical phases: 1) Evaluate: model and characterize agent performance in adaptively generated simulation environments; 2) Challenge: systematically find chaos factors that expose failures; 3) Fitness: balance challenge and feasibility; and 4) Diversity: maintain a population of scenarios that expose the agent to a range of challenges. The work presented here focuses on the Challenge phase and provides preliminary results for our adaptive domain generation paradigm, known as the adversarial architect. The adversarial architect will use an evolutionary-based optimizer to generate diverse scenarios that start simple and become more challenging as agents learn. We demonstrate the efficacy of our approach in the Acrobot domain, where we show the adversarial architect generating an Auto Curricula of inertial parameters that effectively guides the learning agent to resilient and effective policies.