The goal of military training simulators is to portray the realities of combat situations as closely as possible. During combat situations, the performance of military vehicles can sustain progressive degradation induced by a variety of factors that range from enemy fire to crew fatigue. Training simulators should model these degraded states in order to provide military personnel with realistic training environments. Unfortunately, current simulators use less than optimal techniques to model platform degradation. The current techniques are mostly based on a probability of kill (PK). As an example, the performance degradation of a tank is modeled by three states: mobility kill, firepower kill, and catastrophic kill. This model does not leave room for the myriad of degradation conditions that lie somewhere in between these three states, as well as not taking into account other system components, such as communication equipment, nor the degraded performance that can result from human factors unrelated to the state of the equipment such as crew stress and fatigue. Researchers at the Army Materiel Systems Analysis Activity (AMSAA) have developed a new model that proposes a vulnerability and lethality taxonomy (V/LT). This taxonomy serves as a much more realistically metric to describe platform degradation and its resulting consequences. Other researchers, principally Industrial/Organizational (I/O) psychologists, have been employed by the military to determine the influence of human factors in degraded platform behavior.

The purpose of this paper is to examine how to modify the behavior of autonomous intelligent agents (AIPs) given their current degraded state. The proposed method uses the Context-Based Reasoning (CxBR) paradigm to model AIP behavior. The AMSAA V/L taxonomy is incorporated into the model, and performance-degrading human factors are taken into account. To incorporate degraded state behavior into the CxBR paradigm, the current CxBR implementation was modified to incorporate the AIP s degraded state into its reasoning. The modifications changed the CxBR structure by including degraded state knowledge in the AIP fact database, and by altering the reasoning that CxBR uses to choose the appropriate next context. This reasoning is modified by adding weights to each context and functions that calculate these weights. The current context in the proposed implementation is chosen as the context that has received the highest weight. The proposed approach was tested using a small-scale tank warfare scenario with satisfactory results. Future work should implement the concepts presented in this thesis on a larger-scale scenario, and refine implementation details, such as finding optimal functions to calculate the context weights.