Parameter-less Late Acceptance Hill-climbing: Foundations & Applications.

PhD Theses.Stochastic Local Search (SLS) methods have been used to solve complex hard combinatorial problems in a number of elds. Their judicious use of randomization, arguably, simpli es their design to achieve robust algorithm behaviour in domains where little is known. This feature makes them a general purpose approach for tackling complex problems. However, their performance, usually, depends on a number of parameters that should be speci ed by the user. Most of these parameters are search-algorithm related and have little to do with the user's problem. This thesis presents search techniques for combinatorial problems that have fewer parameters while delivering good anytime performance. Their parameters are set automatically by the algorithm itself in an intelligent way, while making sure that they use the entire given time budget to explore the search space with a high probability of avoiding the stagnation in a single basin of attraction. These algorithms are suitable for general practitioners in industry that do not have deep insight into search methodologies and their parameter tuning. Note that, to all intents and purposes, in realworld search problems the aim is to nd a good enough quality solution in a pre-de ned time. In order to achieve this, we use a technique that was originally introduced for automating population sizing in evolutionary algorithms. In an intelligent way, we adapted it to a particular one-point stochastic local search algorithm, namely Late Acceptance Hill-Climbing (LAHC), to eliminate the need to manually specify the value of the sole parameter of this algorithm. We then develop a mathematically sound dynamic cuto time strategy that is able to reliably detect the stagnation point for these search algorithms. We evaluated the suitability and scalability of the proposed methods on a range of classical combinatorial optimization problems as well as a real-world software engineering proble