Visualization for Analyzing Trajectory-Based Metaheuristic Search Algorithms

We call this problem of designing the appropriate metaheuristic problem for a combinatorial (optimization) problem, the metaheuristic tuning problem [1, 3, 7]. Anecdotal evidence suggests that tuning takes a major effort, i.e. [1] states that 90% of the design and testing time can be spent fine-tuning the algorithm. Although it can be easy to come up with a variety of metaheuristics, tuning the metaheuristic implementation is not straightforward. Firstly, the metaheuristics may not be well understood. It might also be applied to problems which may not have been studied. Thus, it may not be clear how to perform tuning. Secondly, the space in which the tuning can operate on is very rich — there are many ways of combining different kinds of metaheuristics each with their own choice of strategies and variations. Furthermore, they each have their own parameters. In this paper, we take a broad view of the metaheuristic tuning problem and understand it to also encompass algorithm design and debugging. Traditionally, the approach for to the tuning problem is either manual experimentation or more automated approaches such as finding the best parameter values [1], best configuration [3], or self-tuning algorithms [2]. In this paper, we take a different approach which takes a human/programmer perspective — how to aid human in solving the tuning problem. Like human-guided search [6], we believe that a cooperative paradigm with the human in the loop can be productive. The difference with human-guided search is that it is concerned with using the human to produce better solutions, while we want to use the human to produce better metaheuristic algorithms. Ultimately, we would like a man-machine cooperation which can help the human to debug, analyze and improve a metaheuristic algorithm for particular problems. Some of the questions which we would like to help answer are: