Training agents in a complex environment

The paper describes an approach to building agents for users of complex data access and management systems for resource and environmental applications. Gathering good examples of this highly specialized and complicated activity is costly and difficult. There is usually only a small set of such good examples available to guide the development of an agent. Consequently, agents are trained, rather than being learned inductively from example sets. In our approach, agents use planning and plan generalization (learning) as their basic mechanism. Plans for yet unseen combinations of goals are created by the merging of plans for individual goals, with the minimum of replanning. An example illustrates merging of existing plans, and shows a simple practical solution to the mutual goal clobbering problem. Plans are built from low-granularity agent commands. The prototype of the system is implemented, and the paper shows a fragment of agent training. The application for this reasoning system addresses the use of planning and of agents to perform forest cover map updates using satellite imagery. To perform this task, a variety of geographical information systems, remote sensing image analysis tools and visualization packages are used