No abstract

No abstract

In what circumstances would you want this AI to make decisions on your behalf?" We have been investigating how to enable a user of an Artificial Intelligence-powered system to answer questions like this through a series of empirical studies, a group of which we summarize here. We began the series by (a) comparing four explanation configurations of saliency explanations and/or reward explanations. From this study we learned that, although some configurations had significant strengths, no one configuration was a clear "winner." This result led us to hypothesize that one reason for the low success rates Explainable AI (XAI) research has in enabling users to create a coherent mental model is that the AI itself does not have a coherent model. This hypothesis led us to (b) build a model-based agent, to compare explaining it with explaining a model-free agent. Our results were encouraging, but we then realized that participants' cognitive energy was being sapped by having to create not only a mental model, but also a process by which to create that mental model. This realization led us to (c) create such a process (which we term After-Action Review for AI or "AAR/AI") for them, integrate it into the explanation environment, and compare participants' success with AAR/AI scaffolding vs without it. Our AAR/AI studies' results showed that AAR/AI participants were more effective assessing the AI than non-AAR/AI participants, with significantly better precision and significantly better recall at finding the AI's reasoning flaws.

Reinforcement learning (RL) agents are commonly evaluated via their expected value over a distribution of test scenarios. Unfortunately, this evaluation approach provides limited evidence for post-deployment generalization beyond the test distribution. In this paper, we address this limitation by extending the recent CheckList testing methodology from natural language processing to planning-based RL. Specifically, we consider testing RL agents that make decisions via online tree search using a learned transition model and value function. The key idea is to improve the assessment of future performance via a CheckList approach for exploring and assessing the agent's inferences during tree search. The approach provides the user with an interface and general query-rule mechanism for identifying potential inference flaws and validating expected inference invariances. We present a user study involving knowledgeable AI researchers using the approach to evaluate an agent trained to play a complex real-time strategy game. The results show the approach is effective in allowing users to identify previously-unknown flaws in the agent's reasoning. In addition, our analysis provides insight into how AI experts use this type of testing approach, which may help improve future instantiations.

Reinforcement learning (RL) agents are commonly evaluated via their expected value over a distribution of test scenarios. Unfortunately, this evaluation approach provides limited evidence for post-deployment generalization beyond the test distribution. In this paper, we address this limitation by extending the recent CHECKLIST testing methodology from natural language processing to planning-based RL. Specifically, we consider testing RL agents that make decisions via online tree search using a learned transition model and value function. The key idea is to improve the assessment of future performance via a CHECKLIST approach for exploring and assessing the agent's inferences during tree search. The approach provides the user with an interface and general queryrule mechanism for identifying potential inference flaws and validating expected inference invariances. We present a user study involving knowledgeable AI researchers using the approach to evaluate an agent trained to play a complex realtime strategy game. The results show the approach is effective in allowing users to identify previously-unknown flaws in the agent's reasoning. In addition, our analysis provides insight into how AI experts use this type of testing approach, which may help improve future instantiations.