Reasoning, learning & coordination: Essays in artificial intelligence and economics
Abstract
This thesis consists of three essays. In the first essay, an integrated model of an artificially intelligent decision-making system is presented which can reason with incomplete knowledge, make revisions in conclusions due to new evidence, and modify the reasoning model itself when it is so warranted. Such a capability is achieved by a synergistic combination of two previously unrelated systems: (1) explanation based learning and (2) truth maintenance systems. The model enhances the paradigm of explanation based learning and reasoning to a world domain characterized by a non-monotonic theory. We explore the problems of generating multiple explanatory models, as a result of having incomplete theory, such as how can different models relate to each other, how to devise an intuitively sound measure of distance in the space of all mental models so that similar mental models can be clustered together, and what principles and criteria to use to arrange competing models in a preference order. The relevance of these concepts are obvious to researchers in distributed artificial intelligence and coordination theory. The focus of the second essay is the process of coordination and cooperation among autonomous units in human systems, in computer systems, and in hybrid organizations. We build an abstract model of the coordination process in which modules interact in an asynchronous fashion. The process of negotiation is described in the framework of asynchronous computation of fixed points. We use concepts from stochastic automata theory and game theory to formally model coordination in hierarchical and matrix organizational structures. Finally, a deterministic process of negotiation is designed in order to explore ideas related to the computational complexity of the process. In the third essay, we show how a Prolog or an expert system can be treated as a computational system in the framework of net theory. Using the Petri net formalism, we model Horn clauses, non-Horn clauses, and expert systems. Once the net representation is established, the collection of analysis techniques associated with net theory can be applied for serving requirements of knowledge engineering. Finally, we show that a major advantage of net theory is the possibility of analyzing parallelism in the inferential process. We specify a formal model that maps the computational model in its net representation to an appropriate architecture which is parallel in nature.
Degree
Ph.D.
Advisors
Whinston, Purdue University.
Subject Area
Management|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.