An architecture for adaptive intelligent systems

TLDR

Adaptive intelligent systems operate in dynamic niches that vary in tasks, resources, conditions, and performance criteria, requiring agents to adapt perceptual strategies, control modes, reasoning tasks, methods, and meta‑control, unlike typical AI agents. The study aims to design and implement an agent architecture that enables comprehensive agents to function effectively in challenging, dynamic niches by constructing explicit control plans for adaptive behavior. The architecture uses a small class hierarchy of AIS niches, exemplified by ICU patient monitoring, and relies on agents dynamically constructing explicit control plans to guide adaptive behavior, as demonstrated by the Guardian experimental agent. The architecture has been successfully applied to build experimental agents across multiple AIS.

Abstract

Our goal is to understand and build comprehensive agents that function effectively in challenging niches. In particular, we identify a class of niches to be occupied by “adaptive intelligent systems (AISs)”. In contrast with niches occupied by typical AI agents, AIS niches present situations that vary dynamically along several key dimensions: different combinations of required tasks, different configurations of available resources, contextual conditions ranging from benign to stressful, and different performance criteria. We present a small class hierarchy of AIS niches that exhibit these dimensions of variability and describe a particular AIS niche, ICU (intensive care unit) patient monitoring, which we use for illustration throughout the paper. To function effectively throughout the range of situations presented by an AIS niche, an agent must be highly adaptive. In contrast with the rather Stereotypie behavior of typical AI agents, an AIS must adapt several key aspects of its behavior to its dynamic situation: its perceptual strategy, its control mode, its choices of reasoning tasks to perform, its choices of reasoning methods for performing chosen tasks; and its meta-control strategy for global coordination of all its behavior. We have designed and implemented an agent architecture that supports all of these different kinds of adaptation by exploiting a single underlying theoretical concept: An agent dynamically constructs explicit control plans to guide its choices among situation-triggered behaviors. The architecture has been used to build experimental agents for several AIS niches. We illustrate the architecture and its support for adaptation with examples from Guardian, an experimental agent for ICU monitoring.