Allocating Tasks in Extreme Teams Paul Scerri ∗ , Alessandro Farinelli + , Steven Okamoto ∗ and Milind Tambe # ∗ Carnegie Mellon University + University of Rome # University of Southern California pscerri@cs.cmu.edu, Alessandro.Farinelli@dis.uniroma1.it, sokamoto@cs.cmu.edu, tambe@usc.edu ABSTRACT Extreme teams, large-scale agent teams operating in dynamic envi- ronments, are on the horizon. Such environments are problematic for current task allocation algorithms due to the lack of locality in agent interactions. We propose a novel distributed task allocation algorithm for extreme teams, called LA-DCOP, that incorporates three key ideas. First, LA-DCOP’s task allocation is based on a dynamically computed minimum capability threshold which uses approximate knowledge of overall task load. Second, LA-DCOP uses tokens to represent tasks and further minimize communica- tion. Third, it creates potential tokens to deal with inter-task con- straints of simultaneous execution. We show that LA-DCOP con- vincingly outperforms competing distributed task allocation algo- rithms while using orders of magnitude fewer messages, allowing a dramatic scale-up in extreme teams, upto a fully distributed, proxy- based team of 200 agents. Varying threshold are seen as a key to outperforming competing distributed algorithms in the domain of simulated disaster rescue. Categories and Subject Descriptors I.2.8 [Artificial Intelligence]: Distributed Artificial Intelligence— Multiagent Systems General Terms Algorithms Keywords Task Allocation, Distributed Constraint Optimization 1. INTRODUCTION Distributed task allocation is a fundamental research challenge in multiagent systems, with recent results reporting significant progress in task allocation in teams[13, 7, 22, 19, 9]. However, a significant large class of practical applications — that we call extreme teams Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. AAMAS’05, July 25-29, 2005, Utrecht, Netherlands. Copyright 2005 ACM 1-59593-094-9/05/0007 ...$5.00. — has emerged, imposing new requirements for task allocation in teams. Extreme teams include mobile sensor or UAV teams[6], robot teams for Mars colonies[5], disaster rescue scenarios, as well as large-scale future integrated manufacturing and service organi- zations (e.g., hospitals)[15]. Extreme teams require team mem- bers, each with limited resources, to act in real-time dynamic en- vironments. More importantly, team members possess overlapping functionality, but differing capabilities to perform different tasks. For instance, in disaster rescue simulations, different fire fighters and paramedics comprise an extreme team; and while fire fighters and paramedics have overlapping functionality to rescue civilians, for a specific rescue task, one set of paramedics may have a higher due to their specific training and current context. The problem of task allocation in teams is one of optimally as- signing tasks in a team plan to agents to maximize overall team utility[12, 22]. Extreme teams emphasize four key constraints on task allocation: (i) domain dynamics may cause tasks to appear and disappear; (ii) agents may perform multiple tasks within resource limits; (iii) many agents have overlapping functionality to perform each task, but with differing levels of capability; and (iv) inter-task constraints (such as simultaneous execution requirements) may be present. This task allocation challenge in extreme teams will be referred to as E-GAP, as it subsumes the generalized assignment problem (GAP), which is NP-complete[21]. The first two constraints in E-GAP above (dynamics and mul- tiple tasks) make approximations necessary, since it is extremely difficult to obtain optimal solutions in a timely fashion. The re- maining two constraints emphasize lack of locality in agent inter- actions, e.g., due to overlapping agent functionality, in assigning a specific task, an agent must potentially consider all other agents (and not a small subset). However, in practical extreme team do- mains agents will frequently possess reasonable estimates of the overall team capabilities or the situation. For example, fire fighter team members may know the number of fire trucks to an order of magnitude, and have (only) a probability distribution on the loca- tions of fires. This imperfect team knowledge is a key property of extreme teams, and provides a valuable way to restrict the search space to good (if suboptimal) solutions. This paper builds on Distributed Constraint Optimization (DCOP)[11, 4] for task allocation, as DCOP offers the key advantages of dis- tributedness, presence of fast/approximate algorithms and a rich representational language which can consider costs/utilities of tasks. Despite these advantages, previous DCOP approaches to task allo- cation suffer from three key weaknesses. First, DCOP algorithms are unable to use imperfect team knowledge to efficiently and ef- fectively allocate tasks. Second, constraints exist between any team