Learning and adoption versus optimization

Ermoliev, Y. (2005). Learning and adoption versus optimization. Lecture Notes in Computer Science 1237 p. 2. 10.1007/3-540-63077-5_20.

Full text not available from this repository.

Abstract

The aim of the talk is to discuss the role of stochastic optimization techniques in designing learning and adaptive processes. Neural and Bayesian networks, path-dependent adaptive urn's processes, automation learning problems and agent based models are considered. A unified general framework of stochastic optimization is proposed enabling us to derive various known and many other adaptive procedures for such different classes of models. We emphasize similarities with natural evolutionary processes, but at the same time we show that this similarity may be misleading when we deal with man-made systems. The "particles" (economic agents, enterprises, countries) of such systems do not follow strong laws like the laws in mechanics and physics (for instance, gravity law). Economic "particles" have flexibility to choose different behavioral patterns (policies, decisions). The uncertainty is a key issue in the modeling of anthropogenic systems and the main methodological challenge is to address related uncertainties explicitly within overall risk-based decision making processes. Purely myopic, trial and error approaches may be expensive, time consuming and even dangerous because of the irreversible nature of decisions. The decisive feature of man-made systems is their ability to anticipate and affect possible future outcomes. Approaches facilitating our ability for making decisions in the presence of uncertainties and related risks are discussed.

Item Type:	Article
Depositing User:	Luke Kirwan
Date Deposited:	04 Aug 2016 12:01
Last Modified:	02 Feb 2022 14:10
URI:	https://pure.iiasa.ac.at/13583

Actions (login required)

View Item