Minerva-DCA: An Intelligent Agent for Ship Damage Control

The decision making task of ship damage control includes addressing problems such as fire spread, floods, smoke, equipment failures, and personnel casualties. It is a challenging and highly stressful domain with a limited provision for real-life training. In response to this need, a multimedia interactive damage control simulator system, called DC-Train 2.0 was recently deployed at a Navy officer training school; it provides officers with an immersive environment for damage control training. This paper describes a component of the DC-Train 2.0 system that provides feedback to the user, called the automated instructor assistant. This assistant is based on a blackboard-based expert system called Minerva-DCA, which is capable of solving damage control scenarios at the "expert" level. Its innovative blackboard architecture facilitates various forms of user assistance, including interactive explanation, advising, and critiquing. In a large exercise involving 500 ship crises scenarios, Minerva-DCA showed a 76% improvement over Navy officers by saving 89 more ships. 1. THE DOMAIN OF SHIP DAMAGE CONTROL The tasks of ship damage control are vital to ship survivability, human life, and operational readiness. Most crises on military and civilian ships could be successfully addressed if handled promptly and properly. Typically the crisis management efforts on a ship are directed by a single person called the Damage Control Assistant (DCA). This person is in charge of maintaining situational awareness, coordinating crisis management crews, and managing other resources. Naturally, crisis management tasks are challenging even for seasoned Navy officers due to the inherent complexity of physical damage, limited resources, information overload, uncertainty, infrequent opportunities for realistic practice, and psychological stress. Studies have shown that the performance could be significantly improved by providing more opportunities for realistic practice [5, 1]. As in many other military domains, real-life training in often infeasible or inadequate due to the high cost and a limited number of possible scenarios. A complement to textbook training is computer simulations [5]. One such damage control simulator, the DC-Train 2.0, is capable of modeling a wide range of realistic scenarios [2]. However, a computer simulation trainer still needs a human instructor to (1) demonstrate a successful scenario solution, (2) provide the novice with instructional advice, (3) observe the novice’s problem-solving and provide a comprehensive critique, and (4) score performance on various scenarios for progress evaluation and comparative analysis purposes. While DCTrain 2.0 is implemented with numerical and knowledge-based simulation techniques, requirements of an automated instructor include (1) an achievement of the level of expertise sufficient to solve arbitrary scenarios in real-time; (2) an ability to observe the novice in real-time, communicate with the novice, and present intelligible feedback in a natural language format. Such functions clearly present an interesting challenge for modern artificial intelligence technology.