Learning to Envision: An Intelligent Agent for Ship Damage Control

This paper describes an Intelligent Agent for real-time crisis decision making, called Minerva-DCA, which improves its performance by compiling the results of a first-principles simulator. The agent is blackboard based and uses envisionment to schedule its actions. This is necessary because the complexity and chaos associated with ship damage control don’t allow the range of necessary behaviors to be easily captured in a ruleset of reasonable size. The first principles envisionment is slow because it involves a simulation of physics of the spread of fires and flooding. In our approach the learning module compiles the knowledge generated by the slow envisionment process and thus allows for real-time performance on subsequent trials. In a large exercise involving 500 ship crises scenarios, Minerva-DCA showed a 76% improvement over Navy officers by saving 89 more ships.