Decision-Theoretic Learning in an Action System

Abstract

The Optimizing Opportunistic Planning System (OOPS) is an action architecture designed to propel a purposeful agent through an environment that is rich with opportunities and threats. It meshes the ethological model of drive hierarchies with connectionist-like perception networks, in order to achieve a smooth integration of perception and action selection. The architecture embodies the idea of computational economy, which yields a uniform procedure for allocating resources to promising perceptual and action hypotheses. This paper discusses the use of tools from decision theory to tune this attention-focusing process, and thereby learn the reliability of, and correlations between, complex features. QOPS's representation scheme applies this learning process to the creation, testing, and identification of cheap features, which predict more expensive internal features, and, ultimately, opportunities in the world.