Learning in large-scale games and cooperative control.

摘要

Many engineering systems can be characterized as a large scale collection of interacting subsystems each having access to local information, making local decisions, having local interactions with neighbors, and seeking to optimize local objectives that may well be in conflict with other subsystems. The analysis and design of such control systems falls under the broader framework of "complex and distributed systems". Other names include "multi-agent control," "cooperative control," "networked control," as well as "team theory" or "swarming." Regardless of the nomenclature, the central challenge remains the same. That is to derive desirable collective behaviors through the design of individual agent control algorithms. The potential benefits of distributed decision architectures include the opportunity for real-time adaptation (or self-organization) and robustness to dynamic uncertainties such as individual component failures, non-stationary environments, and adversarial elements. These benefits come with significant challenges, such as the complexity associated with a potentially large number of interacting agents and the analytical difficulties of dealing with overlapping and partial information.This dissertation focuses on dealing with the distributed nature of decision making and information processing through a non-cooperative game-theoretic formulation. The interactions of a distributed/multi-agent control system are modeled as a non cooperative game among agents with the desired collective behavior being expressed as a Nash equilibrium. In large scale multi-agent systems, agents are inherently limited in both their observational and computational capabilities. Therefore, this dissertation focuses on learning algorithms that can accommodate these limitations while still guaranteeing convergence to a Nash equilibrium. Furthermore, in this dissertation we illustrate a connection between the fields of game theory and cooperative control and develop several suitable learning algorithms for a wide variety of cooperative control problems. This connection establishes a framework for designing and analyzing multi-agent systems. We demonstrate the potential benefits of this framework on several cooperative control problems including dynamic sensor coverage, consensus, and distributing routing over a network, as well as the mathematical puzzle Sudoku.