[Paper Review] HS-CAI: A Hybrid DCOP Algorithm via Combining Search with Context-based Inference

Dingding Chen, Yanchen Deng+3 more|arXiv (Cornell University)|Nov 28, 2019

Constraint Satisfaction and Optimization24 references4 citations

TL;DR

HS-CAI is a novel hybrid DCOP algorithm that integrates tree-based search with iterative context-based inference to produce tighter lower bounds for efficient pruning. By dynamically deriving contexts from the search process and applying a context evaluation mechanism, it reduces inference overhead while outperforming state-of-the-art algorithms in message count, network load, and constraint checks under limited memory budgets.

ABSTRACT

Search and inference are two main strategies for optimally solving Distributed Constraint Optimization Problems (DCOPs). Recently, several algorithms were proposed to combine their advantages. Unfortunately, such algorithms only use an approximated inference as a one-shot preprocessing phase to construct the initial lower bounds which lead to inefficient pruning under the limited memory budget. On the other hand, iterative inference algorithms (e.g., MB-DPOP) perform a context-based complete inference for all possible contexts but suffer from tremendous traffic overheads. In this paper, $(i)$ hybridizing search with context-based inference, we propose a complete algorithm for DCOPs, named {HS-CAI} where the inference utilizes the contexts derived from the search process to establish tight lower bounds while the search uses such bounds for efficient pruning and thereby reduces contexts for the inference. Furthermore, $(ii)$ we introduce a context evaluation mechanism to select the context patterns for the inference to further reduce the overheads incurred by iterative inferences. Finally, $(iii)$ we prove the correctness of our algorithm and the experimental results demonstrate its superiority over the state-of-the-art.

Motivation & Objective

• Address the inefficiency of one-shot approximated inference in existing hybrid DCOP algorithms, which produce weak lower bounds under memory constraints.
• Overcome the high communication overhead of iterative inference algorithms like MB-DPOP by selectively applying inference only to relevant context patterns.
• Develop a complete, optimal DCOP solver that leverages dynamic context derivation from search to improve pruning effectiveness.
• Reduce the computational and communication costs of context-based inference through a context evaluation mechanism that prioritizes high-impact patterns.
• Demonstrate that context-based inference can yield tighter lower bounds than context-free approximations under the same memory budget.

Proposed method

• Integrate a tree-based SBB search with iterative context-based inference, where search provides running contexts for inference to generate tight lower bounds.
• Use context patterns derived from the search process as input for iterative utility propagation in inference, enabling tighter pruning during search.
• Implement a context evaluation mechanism to identify and prioritize the most effective context patterns, reducing the number of required inferences.
• Apply dimension-limited utility propagation in inference to respect memory constraints, similar to ODPOP and MB-DPOP, but with context-aware pruning.
• Dynamically alternate between search and inference phases, where inference results are reused to guide search pruning and reduce the number of contexts to explore.
• Theoretical analysis proves that context-based inference produces lower bounds at least as tight as context-free approximated inference under the same memory budget.

Experimental results

Research questions

• RQ1 Can iterative context-based inference, derived from search contexts, produce tighter lower bounds than one-shot approximated inference in hybrid DCOP algorithms?
• RQ2 How can the number of context-based inferences be minimized without sacrificing pruning effectiveness in memory-constrained environments?
• RQ3 Does combining search with context-aware inference lead to significant reductions in message count, network load, and constraint checks compared to state-of-the-art complete DCOP algorithms?
• RQ4 Can the context evaluation mechanism effectively identify high-impact context patterns to reduce inference overhead while maintaining optimality?
• RQ5 Is the proposed hybrid approach more scalable than existing inference-based (e.g., DPOP, MB-DPOP) and search-based (e.g., PT-FB) algorithms under limited memory budgets?

Key findings

• HS-CAI( k=6 ) solves larger problems than HS-AI( k=6 ) and inference-based algorithms like DPOP and MB-DPOP under low memory budgets, demonstrating superior scalability.
• HS-CAI( k=6 ) incurs fewer messages and less network load than PT-FB and HS-AI, despite the added cost of iterative inference, due to effective pruning from tighter lower bounds.
• HS-CAI( k=6 ) requires fewer NCLOs (number of constraint checks) than HS-AI( k=6 ) when solving problems with 28 agents, indicating faster convergence.
• HS-CAI( k=10 ) produces tighter lower bounds than HS-CAI( k=6 ), but at the cost of higher computation overhead, resulting in more NCLOs but fewer messages.
• HS-CAI outperforms MB-DPOP in both network load and NCLOs on dense configurations, due to selective inference on evaluated context patterns rather than all contexts.
• Theoretical analysis confirms that context-based inference under the same memory budget produces lower bounds that are at least as tight as those from context-free approximated inference.