cuGenOpt: A GPU-Accelerated General-Purpose Metaheuristic Framework for Combinatorial Optimization
New GPU-accelerated framework beats traditional solvers by orders of magnitude on logistics and scheduling problems.
Researcher Yuyang Liu has introduced cuGenOpt, a groundbreaking GPU-accelerated framework designed to tackle complex combinatorial optimization problems that plague industries like logistics, scheduling, and resource allocation. The framework addresses the classic trade-off between generality, performance, and usability by employing a novel 'one block evolves one solution' CUDA architecture. This allows it to massively parallelize the search for optimal solutions across problems encoded as permutations, binary, or integer representations. Benchmarks across three GPU architectures (T4, V100, A800) show it outperforms general Mixed-Integer Programming (MIP) solvers by orders of magnitude and achieves competitive results against specialized solvers on problems with up to 150 variables.
At its core, cuGenOpt features a two-level adaptive operator selection mechanism and hardware-aware resource management to maximize GPU efficiency. Its extensibility is a key strength, allowing domain experts to inject custom, problem-specific CUDA search operators. For usability, it offers a pure-Python API via a Just-In-Time (JIT) compilation pipeline. Most notably, it includes an LLM-based modeling assistant that can translate a natural language description of a problem—like 'schedule deliveries for 50 trucks across a city'—directly into working solver code. In tests, framework optimizations cumulatively reduced the solution gap for a complex PCB routing problem (pcb442) from 36% to just 4.73%.
- Uses a 'one block evolves one solution' CUDA architecture to achieve 75-81% higher throughput for vehicle routing problems.
- Integrates an LLM assistant that converts natural-language problem descriptions directly into executable optimization code.
- Demonstrated a 4.73% solution gap on the challenging TSP-442 problem within just 30 seconds on modern GPUs.
Why It Matters
Dramatically accelerates solving real-world logistics and scheduling puzzles, making advanced optimization accessible without deep expertise.