Parallel Branch-and-Bound Algorithms for Combinatorial Optimization

Branch-and-Bound represents a cornerstone methodology for solving discrete optimisation problems by systematically exploring a search tree of candidate solutions. Each node in the tree corresponds to a partial assignment of decision variables, and a bounding function estimates the best possible outcome that can be obtained from that node. If the bound is inferior to an incumbent solution, the branch is pruned, thereby reducing the search space. While the sequential form of the algorithm has proven effective on moderate-scale instances, modern combinatorial problems often demand capacities beyond single-processor throughput. Parallel Branch-and-Bound approaches distribute subtrees or subproblems across multiple processing units, aiming to exploit concurrency while preserving pruning efficiency. Key challenges include irregular workload distribution due to highly unbalanced tree structures, communication overheads for sharing global incumbents, and redundant exploration of overlapping regions of the search tree. Contemporary implementations harness diverse parallel paradigms, from master–worker and work-stealing frameworks to data-parallel platforms such as MapReduce. High-performance computing environments have enabled solutions to large-scale mixed-integer programmes, complex scheduling tasks and combinatorial designs in computational biology, demonstrating order-of-magnitude speedups. Advances in dynamic load balancing, decentralised incumbent dissemination and hybrid CPU–GPU execution models continue to bridge the gap between theoretical scalability and practical efficiency, underscoring the global significance of parallel Branch-and-Bound in resource allocation, logistics and beyond.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

A novel AND/OR Branch-and-Bound scheme deployed on computational grids has showcased near-optimal parallel speedups for a wide range of combinatorial benchmarks. By organising the search space into an AND/OR graph, the method minimises redundant exploration and employs machine-learning-guided load balancing to redistribute work between nodes, overcoming synchronisation bottlenecks. Experimental campaigns on hundreds of CPUs solved instances previously deemed intractable, although analysis reveals that communication latency and search-space overlap still cap scalability in extreme configurations.

Another development leverages the MapReduce paradigm to tackle job-shop scheduling via parallel Branch-and-Bound. Subproblems are generated by partitioning the schedule horizon and dispatched as independent map tasks. Reducer functions integrate local incumbents and perform global pruning, dramatically cutting communication volume compared to traditional master–worker schemes. This approach has delivered substantial speedups on cloud infrastructures, facilitating real-time decision support in manufacturing scenarios.

Efforts to exploit large-scale supercomputers have also borne fruit in mixed-integer programming. By distributing nodes of the Branch-and-Bound tree across over a million CPU cores, researchers achieved unprecedented solution times on single hard instances. A hybrid strategy combining decentralised incumbent sharing with asynchronous work-stealing mitigates load imbalance, while low-latency interconnects and tailored bounding functions preserve pruning efficacy at scale. These results highlight the potential and limits of exascale Branch-and-Bound for industrial-strength optimisation.

Parallel Branch-and-Bound Algorithms for Combinatorial Optimization publication trend

The graph below shows the total number of articles in parallel branch-and-bound algorithms for combinatorial optimization across all publications each year (not limited to Nature Index journals).

Technical terms

Branch-and-Bound: A tree-based algorithm that systematically partitions and prunes the search space to solve optimisation problems exactly.

Combinatorial optimisation: The task of finding an optimal arrangement or selection from a finite, often discrete, set of possibilities under given constraints.

Search tree: A hierarchical representation of partial solutions where each branch corresponds to a decision or assignment step.

Bounding function: A heuristic or calculation that provides an upper or lower limit on the optimal solution within a subtree, used to prune unpromising branches.

Load balancing: The dynamic allocation or migration of computational tasks across processors to ensure even utilisation and minimise idle time.

Pruning: The elimination of branches of the search tree that cannot yield better solutions than the best found so far, based on bounding information.

References

1. AND/OR Branch-and-Bound on a Computational Grid. Journal of Artificial Intelligence Research (2017).
2. Parallel Branch and Bound Algorithm for Product Testing Job Scheduling Problems using MapReduce. International Journal of Machine Learning and Computing (2020).
3. Harnessing over a Million CPU Cores to Solve a Single Hard Mixed Integer Programming Problem on a Supercomputer. Proceedings of International Conference on Artificial Life and Robotics (2018).