Learning strategies for dynamic decision problems using artificial neural networks

The architecture of an AI (artificial intelligence) system for learning strategies in complex domains is presented. SMALL (strategy acquisition by meta-level learning) is a system architecture that embodies the principles of modular knowledge-level design and phased training in order to learn strategies in a flexible yet efficient manner. A class of difficult decision problems is identified. It is shown that specific bodies of knowledge can be used to counter the specific aspects or difficulty. The authors´ approach is illustrated by a connectionist implementation of the knowledge modules. This approach can be used for learning load-balancing strategies in loosely coupled multiprocessors