Optimization and coordination of multi-agent systems using principled negotiation

Principled negotiation coordinates the actions of agents with different interests, allowing distributed optimization. In principled negotiation, agents search for and propose options for mutual gain. If the other agents agree to the proposal, it is implemented. Under certain conditions, an agent can search for options for individual gain without impacting other agents. In these cases, the agent can negotiate with a coordinator, rather than obtain agreement from all other agents. The tenets of principled negotiation are outlined and stated mathematically. Two problems representing air traffic operations are formulated to test the performance of principled negotiation. The first, based on keeping separation between aircraft, has no coupling between the agent actions if certain requirements are met. Principled negotiation allows the agents to achieve a solution as good as that achieved by a centralized controller with perfect knowledge. The second problem, based on negotiating arrival slots, is highly coupled, constraining each agent's available set of actions. Principled negotiation allows agents to search options that would not be available otherwise, improving the utility function of all agents. Principled negotiation can be quickly introduced into air traffic operations. INTRODUCTION The ground-based air traffic control (ATC) system was created to insure the safety of flights operating in controlled airspace. Aircraft are separated by a combination of procedures and tactical maneuvering instructions. As air traffic has grown, the ATC system has increasingly depended on computer systems. Computers now not only process radar and flight plan data, but also help controllers to manage flow, avert conflicts, and maneuver traffic in terminal areas [1,2]. Today's ATC system has many problems that are characteristic of traditional control systems for largescale industrial systems [3]. To manage the growing amount of air traffic, ATC computer systems arc becoming more complex, increasing expense and making new systems more difficult to introduce. In * Research Assistant ** Professor, Associate Fellow AIAA Copyright © 1996 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved. addition, the aircraft/airspace system (AAS) is not responsive to the desires of users (aircraft and operators). The procedures and actions of the ATC system prevent users from dynamically optimizing their operations and causes many hours of delays. Distributed artificial intelligence (DAT) deals with small, simple systems working cooperatively to better control large-scale systems. In multi-agent systems (MAS), each agent has its own goals, and it must anticipate the actions of other agents and coordinate actions to meet these goals. The AAS is a MAS. It is a collection of agents, each with its own goals and interests. Agents include aircraft, operators, and traffic management agents (TrMAs, a generic term for any air traffic control unit). Each agent makes decisions and takes actions that affect the air traffic process. Their actions interact because aircraft must stay safely separated. The ATC system coordinates the actions of agents because, until recently, only the ATC system had sufficient data (on traffic, flight plans, and the weather) and sufficient computing power to analyze the situation. Now, airlines and aircraft also have powerful computer systems, and they can access large amounts of data from their own sensors and through high-bandwidth communications. They are also capable of making declarative decisions regarding the traffic situation. Steeb et al [4] studied whether aircraft alone could resolve conflicts. When a conflict arose, the affected aircraft used a variety of criteria to determine which aircraft was best-suited to formulate a resolution plan. The chosen aircraft then calculated the plan and transmitted it to the other aircraft. This was a centralized control system, but the air traffic process was broken down into distributed conflict areas each with a controlling aircraft. Davis and Smith used a contract net approach to assign surveillance tasks for particular areas to individual aircraft [5], A manager divided the task and issued a request for bids. The agents then sent in bids, and the manager selected the successful bidders. Levy and Rosenschein distributed the coordination function using game theory [6]. In the Pursuit Problem, each pursuer first evaluated the solution of the local game to calculate the total payoff received by all the agents from their combined actions. Each agent then solved the global game to establish its share of the