Distributed Generalized Assignment Problem (D-GAP) is very popular in scalable multi-agent systems. However, existing algorithms are not effective or efficient in large scale or highly dynamic domains due to the limited communication and computation resource. In this paper, we present a novel approach to address this issue. To reduce communication, we propose a decentralized model for agents to jointly search for optimized solutions. Considering the complexity of D-GAP in a massive multi-agent system, agents cannot perform the optimal search based on their local views, we propose a heuristic algorithm. By inferring knowledge from their previous communicated searches, agents are able to predict how to deploy their future similar searches more efficiently. If an agent can solve some parts of D-GAP well, similar searches will be sent to it. By taking the advantage of the accumulation effect to agents' local knowledge, agents can independently make simple decisions with highly reliable performance and limited communication overheads. Finally, we present a simulation to demonstrate the feasibility and efficiency of our algorithm.
[1]
Milind Tambe,et al.
Allocating tasks in extreme teams
,
2005,
AAMAS '05.
[2]
Makoto Yokoo,et al.
The Distributed Constraint Satisfaction Problem: Formalization and Algorithms
,
1998,
IEEE Trans. Knowl. Data Eng..
[3]
Makoto Yokoo,et al.
An asynchronous complete method for distributed constraint optimization
,
2003,
AAMAS '03.
[4]
Yang Xu,et al.
An integrated token-based algorithm for scalable coordination
,
2005,
AAMAS '05.
[5]
Jonathan P. How,et al.
Cooperative path planning for multiple UAVs in dynamic and uncertain environments
,
2002,
Proceedings of the 41st IEEE Conference on Decision and Control, 2002..
[6]
Milind Tambe,et al.
Token approach for role allocation in extreme teams: analysis and experimental evaluation
,
2004,
13th IEEE International Workshops on Enabling Technologies: Infrastructure for Collaborative Enterprises.