Título: Energy Efficiency Metrics for Reinforcement Learning on the Travelling Salesman Problem: An Experimental Approach in Green Computing
Autores: Celestino Simon, Samara Oliveira Silva Santos, Thalita Nazare, André Luiz Carvalho Ottoni, Danton Diego Ferreira & Dr Erivelto n Nepomuceno
Resumo: Artificial intelligence models continue to grow in complexity, but so does their energy demand—an issue gaining attention as these models expand into energy-sensitive domains such as Reinforcement Learning. This paper examines the energy usage of Q-learning in solving the Travelling Salesman Problem. To carry out this analysis, both hardware and software tools were used to measure energy: a PMD-USB power meter for direct readings and CodeCarbon for software-based estimation. By comparing a standard version of Q-learning with one that includes ϵ-decay, the results showed clear improvements. The optimised version completed tasks faster and used noticeably less energy in some cases, up to 68% less. These results suggest that with minor adjustments, it is possible to make Machine Learning models more energy-efficient. The approach used here could be beneficial in other areas that rely on AI but require energy efficiency, such as logistics or embedded systems. Beyond its core findings, the study also underscores the importance of validating software-based estimates with real-time hardware monitoring. The dual platform comparison provides a methodological basis for benchmarking future RL models under energy constraints. This contribution supports a growing interest in evaluating Machine Learning not only by performance, but also by sustainability.
Palavras-chave: Artificial Intelligence (AI); Green Machine Learning; Energy Efficiency; Green Computing; Sustainable AI; Travelling Salesman Problem(TSP); Reinforcement Learning (RL); Power Measurement.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191175
Artigo em PDF: CBIC_2025_paper1191175.pdf
Arquivo BibTeX:
CBIC_2025_1191175.bib