Date of Award
2026
Degree Type
Thesis
Degree Name
Master of Science in Electrical Engineering (MSEE)
Department
Electrical, Computer, and Biomedical Engineering
First Advisor
Manbir Sodhi
Second Advisor
Resit Sendag
Abstract
This thesis presents two complementary contributions to the field of GPU-accelerated evolutionary metaheuristics for combinatorial optimization, organized in manuscript format.
The first manuscript, “BrkgaCuda 3.0: A Redesigned Multi-GPU Framework for Biased Random-Key Genetic Algorithms,” presents a ground-up architectural redesign of BrkgaCuda 2.0 that enables a true multi-GPU island model for Biased Random-Key Genetic Algorithms (BRKGA). The BRKGA island model evolves multiple semi-independent populations that periodically exchange elite solutions, a structure that maps naturally to multi-GPU parallelism; however, BrkgaCuda 2.0 is confined to a single GPU. BrkgaCuda 3.0 introduces an IslandManager that distributes populations across any number of GPUs, with multiple populations per GPU supported via concurrent CUDA streams and a delayed asynchronous peer-to-peer migration protocol. The design also replaces all device-wide synchronization with CUDA event-based stream dependencies and pre-allocates all buffers to eliminate per-generation memory overhead. Scaling from 1 to 8 GPUs on TSP nu3496 yields 34% better solutions with only 3% additional wall-clock time. The cleaner architecture also improves single-GPU performance substantially: across 132 benchmark instances (TSP, SCP, CVRP) with 2,640 total runs, BrkgaCuda 3.0 achieves a 4.6×wall-clock speedup over v2.0 with equivalent solution quality. Profiling confirms that this gain comes from eliminating synchronization overhead rather than from faster kernels, a finding that generalizes to any iterative GPU framework.
The second manuscript, “LLM-Driven Configuration of Genetic Algorithms for Constraint Rich Optimization Problems,” presents a framework that combines large language models (LLMs) for problem understanding with a structured BRKGA configurator for algorithm realization, allowing users to describe optimization problems in natural language and receive executable GPU-accelerated GA implementations. We compare two workflows: LLM+Config, where the LLM extracts problem semantics and a modular BRKGA framework provides the evolutionary architecture with ALNS warm-start and local search; and LLM-Only, where the LLM generates a complete standalone GA. LLM+Config outperforms or matches LLM-Only across all problem types, with the advantage growing as constraint complexity increases. On standard TSP, LLM+Config achieves optimal solutions on 8 of 14 instances. On VRP-RPD, LLM+Config wins all 13 instances against LLM-Only, with 51-329% advantage on large instances. GPU-parallel decoding achieves 25-36×speedup over CPU.
Together, these manuscripts advance both the computational infrastructure and the usability of GPU-accelerated genetic algorithms for combinatorial optimization.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Harishjitu Saseendran, Fnu, "GPU-ACCELERATED BIASED RANDOM-KEY GENETIC ALGORITHMS: FRAMEWORK OPTIMIZATION AND LLM-DRIVEN CONFIGURATION" (2026). Open Access Master's Theses. Paper 2728.
https://digitalcommons.uri.edu/theses/2728