Sergei Nirenburg, Professor, Cognitive Science - Modeling Wisely Lazy Agents

Sergei Nirenburg, Professor, Cognitive Science - Modeling Wisely Lazy Agents

Human language processing, reasoning, decision making and action are influenced by a broad spectrum of factors, such as, to name a few, rational problem solving ability, beliefs, skills, wisdom, erudition, character traits, subconscious biases, emotional abilities, social adeptness, ethical backbone and esthetic sensibility. It is natural to expect that modeling human behavior with the help of artificial intelligent agents will address many, if not all, of these factors in an integrated computational system. This, however, is a tall order: the overall problem becomes very complex very fast.  Following the classical tenets of the scientific method – using abstraction and decomposing a problem into a set of constituent parts – most researchers are pursuing more or less shallow descriptions of more or less narrow-scope partial tasks. This approach permits incremental progress but does not necessarily advance the overall task of modeling human behavior. Eventually, progress on the latter task will require configuring and continuously testing non-toy integrated systems comprising many knowledge and processing components. So, there is tension between what must be done eventually and what can be done in the immediate future.

Human decision makers (subconsciously) experience a similar tension – they must function well and at the same time function with quite limited resources (of time, energy, skills and knowledge). Judicious use of resources seems to be the crucial metalevel ability that distinguishes successful agents. These agents are working hard to be as lazy as they can – to avoid high cognitive loads and waste of resources. A useful concept in this regard is the principle of least effort. In this talk I will discuss some possible heuristics to support a computational implementation of this principle, which, I believe, transcends the well-known notions (due to Simon) of bounded rationality and satisficing.


Download paper here.