DRAFT VERSION – Final version to be presented at 2001 AAAI Fall Symposium Marbles: A Family of Cooperative Negotiation Schemes for Real-Time Fault-Tolerant Distributed Resource Allocation Martin Frank, Alejandro Bugacov, Jinbo Chen, Gordon Dakin, Pedro Szekely, Bob Neches Information Sciences Institute of the University of Southern California 4676 Admiralty Way Marina del Rey, California 90292 {frank,bugacov,jinbo}@isi.edu, gad@crystaliz.com , {szekely,rneches}@isi.edu Abstract Marbles schemes are a family of cooperative and adaptive algorithms for distributed resource allocation problems. Long-term goals for these schemes emphasize fault- tolerance and real-time performance in which a good timely solution is preferable to an optimal but too late solution. This paper reports work in progress where we compare the performance and analyze characteristics of different Marbles schemes and centralized solvers working on large scale and complex resource allocation problems. Introduction – the Marbles Vision Recent advances in miniaturization and robotics have led to interest in research on the “autonomous agents” that make up a team of, say, robotic soccer players or Unmanned Combat Air Vehicles (UCAVs). These agents can act individually but are better off coordinating with their peers. A subset of this problem is distributed real-time resource allocation – deciding under time pressure which soccer player will take the final shot on the goal, or which UCAVs will neutralize a newly discovered enemy threat. There is a spectrum of approaches for distributed real-time resource allocation, ranging from no communication at all (physics-based approach: agents observe each others’ behavior but do not explicitly communicate, much like a wolf pack closing in on prey) to communication of the full rationale of behavior (argumentation-based approach: agents back up requests to others by an argument of why they should grant it). In this continuum, our Marbles schemes and all other “market-inspired” approaches fall in-between: The complexity of the messages they exchange is smaller than in the argumentation approach, yet the prices set by demand and supply can often communicate rationale in an alternative, more compact fashion, and – potentially – steer the group of agents to sensible behavior via “the invisible hand of the market”, requiring significantly less communication. Copyright © 2000, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. Market-based negotiation schemes obviously use more messages than a purely physics-based approach, but can also explore a more complex set of alternatives; and we believe that e.g. any heterogeneous swarm of UCAVs is better off having a pre-made master attack plan that can be adjusted in transit before reaching the targets (which invariably requires messaging). This is especially true if multiple UCAVs (e.g two attack ones, a supervisory one, and a laser designator one) have to collaborate to attack a single target. External Marbles Scheme Properties “Marbles” schemes 1 are a family of resource allocation algorithms that are characterized by the following properties: Distributed. Each task only knows about its local requirements, and communicates with potential resources for those requirements exclusively through messages. Hence, each task and each resource can – but does not have to – be located on a different machine. Cooperative. Marbles schemes are not designed to tolerate malicious participants, which distinguishes our research from work on e.g. electronic commerce and automated auctions; we believe that security against external attack of cooperative negotiation schemes is best located at a lower level (such as the message transport and encryption level). The cooperative nature of the negotiation means that tasks participating in resource auctions can voluntarily “kill themselves” if they conclude that they are unlikely to succeed yet prevent others from succeeding by bidding up resource prices; the distributed algorithms for this “altruistic task suicide” phase further distinguish our work from work on competitive auctions. Adaptive. A Marbles scheme can adapt a current partial solution to a new situation rather than having to re-compute the new solution from scratch. This makes them applicable in cases where “the world can’t stop while a solver computes a solution for everyone”, that is, in cases where 1 The name is not an acronym. Bob Neches likened the agent behavior to “kids trading marbles” in an early design discussion, and the name stuck.