Human behavior representations (HBRs) are an essential component of simulation-based training. Historically, these HBRs involve significant cost to encode expert knowledge for specific uses. One of the largest development costs is re-engineering due to failure, where failures stem from incorrect, incomplete, obsolete, or inconsistent information.

Previous work has demonstrated that it is possible for HBRs to overcome a significant number of their own failures by providing them with limited meta-awareness through self-monitoring, error-detection, and failure recovery. A limitation of those approaches has been their reliance on random selection among possible actions. Although self-correcting systems utilizing random selection are theoretically capable of eventually searching the entire recovery space, this is undesirable in real-time applications. Pruning the search space to decrease the number of options that must be attempted before achieving success can reduce the impact on real-time applications as well as reducing the incidence of recovery-induced failure. However, it can be developmentally intensive to manually annotate world states with suggested recovery actions.

This paper describes a complementary methodology that improves the quality of recovery actions without significantly increasing development cost. Specifically, this paper examines constraints on failure recovery through the use of causal models and learning. These constraints drastically reduce the search possibilities over exhaustive recovery techniques. First, HBRs are enhanced with internal representations of causal models describing recovery domain spaces. These models guide the reasoning process to the selection of appropriate, relevant recovery actions. Once a recovery method is attempted, the self-monitoring capability of the agents is utilized to measure the efficacy of the selected recovery actions. Finally, a learning mechanis m, which will be described in detail, is used to prefer future selection of recovery actions in similar circumstances.