[Paper Review] Distributed Heuristic Forward Search for Multi-Agent Systems

Raz Nissim, Ronen I. Brafman|arXiv (Cornell University)|Jun 25, 2013

Logic, Reasoning, and KnowledgeComputer Science34 references16 citations

TL;DR

This paper presents MAD-A*, a distributed heuristic forward search algorithm for multi-agent systems that enables efficient, privacy-preserving planning by distributing computation across agents while respecting private information. It outperforms both centralized and existing distributed planners by exploiting problem structure to reduce search space and avoid information leaks.

ABSTRACT

This paper describes a number of distributed forward search algorithms for solving multi-agent planning problems. We introduce a distributed formulation of non-optimal forward search, as well as an optimal version, MAD-A*. Our algorithms exploit the structure of multi-agent problems to not only distribute the work efficiently among different agents, but also to remove symmetries and reduce the overall workload. The algorithms ensure that private information is not shared among agents, yet computation is still efficient -- outperforming current state-of-the-art distributed planners, and in some cases even centralized search -- despite the fact that each agent has access only to partial information.

Motivation & Objective

• To develop a distributed planning algorithm that respects the inherent structure of multi-agent systems, where each agent controls private actions and information.
• To ensure privacy by preventing agents from sharing private state or action details, even during search.
• To achieve performance comparable to or better than centralized search, despite limited local information.
• To provide a scalable and robust alternative to centralized planning, especially in systems where a trusted party does not exist.
• To lay the foundation for future work on quantifying privacy loss and enhancing privacy guarantees in distributed search.

Proposed method

• The algorithm uses a distributed formulation of forward search, where each agent maintains its own local search space based on its private actions and preconditions.
• Agents coordinate via public actions and intermediate public states, exchanging only state information necessary for planning.
• A heuristic function is used to guide search, with agents computing local cost estimates and propagating upper bounds on action costs.
• The algorithm employs pruning techniques derived from distributed search to reduce redundant exploration and improve efficiency.
• Private actions are used to satisfy preconditions of public actions, and agents compute upper bounds on the cost of applying public actions without revealing private details.
• The method supports both optimal (MAD-A*) and non-optimal search variants, with the latter avoiding disclosure of g-values to prevent information leakage.

Experimental results

Research questions

• RQ1 Can a distributed forward search algorithm achieve performance comparable to centralized search while preserving agent privacy in multi-agent planning?
• RQ2 How can private information be protected during distributed planning when agents have limited access to each other’s actions and states?
• RQ3 What mechanisms can be used to compute cost bounds for public actions without revealing private action sequences or internal states?
• RQ4 To what extent can distributed search exploit problem structure to reduce the overall search space and improve efficiency?
• RQ5 How can privacy-preserving techniques be integrated into distributed search without significantly degrading performance or solution quality?

Key findings

• MAD-A* outperforms current state-of-the-art distributed planners and, in some cases, even centralized search algorithms, despite each agent having only partial information.
• The algorithm effectively reduces search space by exploiting the distributed structure of multi-agent problems and removing symmetries through coordinated action application.
• Private information is preserved throughout the search process, as agents do not share internal states, action sequences, or preconditions.
• The method enables efficient computation of upper bounds on the cost of applying public actions, even when private actions are involved.
• The work reveals a novel pruning technique for centralized planning that emerges from distributed search, demonstrating broader applicability beyond distributed systems.
• Empirical results suggest that privacy leaks—such as inference of private state models—are possible but limited, especially in problems solved quickly with few expanded nodes.