Exploring Tabu Tenure Policies with Machine Learning

Tabu search is a well-known local search-based metaheuristic, widely used for tackling complex combinatorial optimization problems. As with other metaheuristics, its performance is sensitive to parameter configurations, requiring careful tuning. Among the critical parameters of tabu search is the tabu tenure. This study aims to identify key search attributes and instance characteristics that can help establish comprehensive guidelines for a robust tabu tenure policy. First, a review different tabu tenure policies is provided. Next, critical baselines to understand the fundamental relationship between tabu tenure settings and solution quality are established. We verified that generalizable parameter selection rules provide value when implementing metaheuristic frameworks, specifically showing that a more robust tabu tenure policy can be achieved by considering whether a move is improving or non-improving. Finally, we explore the integration of machine learning techniques that exploits both dynamic search attributes and static instance characteristics to obtain effective and robust tabu tenure policies. A statistical analysis confirms that the integration of machine learning yields statistically significant performance gains, achieving a mean improvement of 12.23 (standard deviation 137.25, $n=$ 10,000 observations) when compared to a standard randomized tabu tenure selection (p-value < 0.001). While the integration of machine learning introduces additional computational overhead, it may be justified in scenarios where heuristics are repeatedly applied to structurally similar problem instances, and even small improvements in solution quality can accumulate to large overall gains. Nonetheless, our methods have limitations. The influence of the tabu tenure parameter is difficult to detect in real time during the search process, complicating the reliable identification of when and how tenure adjustments impact search performance. Additionally, the proposed policies exhibit similar performance on the chosen instances, further complicating the evaluation and differentiation of policy effectiveness.