Portland State University. Systems Science Ph. D. Program.
George G. Lendaris
Date of Award
Doctor of Philosophy (Ph.D.) in Systems Science
3, xi, 256 leaves: ill. 28 cm.
Economics -- Mathematical models
Economic systems are distributed in the sense that economic agents make decisions without any central control. Prices, quantities, wealth, and market structure emerge from the interaction of agents acting in their own self interest. The concepts and language of systems science are used to define economic systems in a manner that captures and articulates the distributed nature of economic systems. Further, the systems definition permits multiple views of the economic system, and in addition, allows the agents to "step outside" the system in order to study it. Economic systems are defined in such a way that it is feasible to construct artificial economic systems, and in particular, ones that are composed of self-interested agents that operate according to principles that are prescribed by the researcher. An artificial economic system was actually constructed and tested in a computer environment. The model was verified with reference to several theoretical models such as static and adaptive expectations. The system constructed allows up to 1000 agents to interact without any central control. A computer "blackboard system" is used as the architecture for providing common information to the agents in the artificial economic system. The blackboard design successfully allows complex agents to compete and trade in an artificial economic system created by the researcher. Prices, quantities, wealth, and market structure emerge naturally in the artificial economy that depend on the characteristics and prescribed strategies of the agents in the system. After a transition period, the trading frequently produces price and quantity time series that have the characteristics of a random walk, a condition that is well known in real world markets. Three classes of producer agents were used in these artificial economic systems: optimizing agents that incorporate neural networks, satisficing agents that incorporate very simple rule-based approaches, and Stackelberg agents that have knowledge about the consumers in the system, but do not have knowledge about their competitor's strategies or intentions. Neural networks are used to model the behavior and strategies of economic agents that can be said to learn, i.e., those agents that develop general principles for adapting to changing market conditions that transfer across markets. The focus of this research was on the producers in the system. The consumption side of the economic system was represented by a set of simple consumers. An important result emerging from this research is that at least one agent out of four in these experiments with accurate knowledge about market demand increases the wealth of the system as a whole. Markets containing a single Stackelberg or neural agent produced far more wealth than markets composed only of satisficing agents. However, the agents with knowledge do not necessarily capture the highest share of the wealth. The success of individual agents depends on the agent's trading strategy, as expected, and in addition depends on the combination of agents in the system. Certain strategies appeared to be flexible while others were brittle, and were easily foiled by changing the agents in the market, or by changing the market conditions. Earlier studies attempted to use neural networks to simulate an entire economic system, but were rejected because the organizing principles of the two systems are not analogous. Additionally, neural networks were successfully tested for solving various economics problems that were not related to the simulation of economic systems. Neural networks were found to effectively solve problems with missing and redundant data that are not directly solvable with well known methods such as least squares.
Perry, Stanley Foster, "Distributed Economic Systems with Agents that Learn" (1992). Dissertations and Theses. Paper 1271.