Victor R. Lesser, Emeritus Distinguished Professor, UMass Amherst


Victor R. Lesser, Emeritus Distinguished Professor, UMass Amherst

Sage 4101

October 14, 2015 12:00 PM - 1:00 PM

Abstract: I have been building self-awareness into AI systems for over 40 years through the basic paradigm of 1) detection of unexpected behavior, 2) diagnosis to understand the violated assumptions that are responsible for the unexpected behavior, 3) adjusting parameters based on revised assumptions and 4) monitoring subsequent behavior for whether behavior is now in line with new expectations. I will share my experiences with you on the different ways that this type of self-awareness can be built into a computational system and why it is an important component to incorporate in the next generation of complex AI systems. Complex AI systems have many parameters that guide their internal problem solving and these parameters need to be set appropriately in order for the system to function effectively. The setting of these parameters is based on the designer’s understanding of the characteristics of how problem solving works and the environmental conditions in which problem solving occurs. These understandings can be thought of as problem-solving assumptions; however, these assumptions may not be true for every problem-solving instance the system encounters or may become incorrect as the environment changes. Self-aware systems are able to adjust these parameters in light of both changing environmental and internal processing characteristics. This is different from what is commonly considered on-line learning, where the adaptation is incremental (gradual) and implicit; instead, adaptation in a self-aware system is more explicit in response to emerging conditions and occurs at specific points in time. However I see a system needing both types of adaptation and, in fact, these two types of adaptation can co-exist and mutually interact. I will conclude with a discussion on a recent multi-agent system that my research group has built that contains both types. 




Add to calendar