Steven Halim

Viz: a visual analysis suite for explaining local search behavior

NP-hard combinatorial optimization problems are common in real life. Due to their intractability, local search algorithms are often used to solve such problems. Since these algorithms are heuristic-based, it is hard to understand how to improve or tune them. We propose an interactive visualization tool, VIZ, meant for understanding the behavior of local search. VIZ uses animation of abstract search trajectories with other visualizations which are also animated in a VCR-like fashion to graphically playback the algorithm behavior. It combines generic visualizations applicable on arbitrary algorithms with algorithm and problem specific visualizations. We use a variety of techniques such as alpha blending to reduce visual clutter and to smooth animation, highlights and shading, automatically generated index points for playback, and visual comparison of two algorithms. The use of multiple viewpoints can be an effective way of understanding search behavior and highlight algorithm behavior which might otherwise be hidden.

Engineering Stochastic Local Search for the Low Autocorrelation Binary Sequence Problem

This paper engineers a new state-of-the-art Stochastic Local Search (SLS) for the Low Autocorrelation Binary Sequence (LABS) problem. The new SLS solver is obtained with white-box visualization to get insights on how an SLS can be effective for LABS; implementation improvements; and black-box parameter tuning.

Tuning Tabu Search Strategies Via Visual Diagnosis

While designing working metaheuristics can be straightforward, tuning them to solve the underlying combinatorial optimization problem well can be tricky. Several tuning methods have been proposed but they do not address the new aspect of our proposed classification of the metaheuristic tuning problem: tuning search strategies. We propose a tuning methodology based on Visual Diagnosis and a generic tool called Visualizer for Metaheuristics Development Framework(V-MDF) to address specifically the problem of tuning search (particularly Tabu Search) strategies. Under V-MDF, we propose the use of a Distance Radar visualizer where the human and computer can collaborate to diagnose the occurrence of negative incidents along the search trajectory on a set of training instances, and to perform remedial actions on the fly. Through capturing and observing the outcomes of actions in a Rule-Base, the user can then decide how to tune the search strategy effectively for subsequent use.

An integrated white+black box approach for designing and tuning stochastic local search

Stochastic Local Search (SLS) is a simple and effective paradigm for attacking a variety of Combinatorial (Optimization) Problems (COP). However, it is often non-trivial to get good results from an SLS; the designer of an SLS needs to undertake a laborious and ad-hoc algorithm tuning and re-design process for a particular COP. There are two general approaches. Black-box approach treats the SLS as a black-box in tuning the SLS parameters. White-box approach takes advantage of humans to observe the SLS in the tuning and SLS re-design. In this paper, we develop an integrated white+black box approach with extensive use of visualization (white-box) and factorial design (black-box) for tuning, and more importantly, for designing arbitrary SLS algorithms. Our integrated approach combines the strengths of white-box and black-box approaches and produces better results than either alone. We demonstrate an effective tool using the integrated white+black box approach to design and tune variants of Robust Tabu Search (Ro-TS) for Quadratic Assignment Problem (QAP).

A software framework for fast prototyping of meta-heuristics hybridization

Hybrids of meta-heuristics have been shown to be more effective and adaptable than their parents in solving combinatorial optimization problems. However, hybridized schemes are also more tedious to implement due to their increased complexity. We address this problem by proposing the meta-heuristics development framework (MDF). In addition to being a framework that promotes software reuse to reduce developmental effort, the key strength of MDF lies in its ability to model meta-heuristics using a “request, sense and response” schema, which decomposes algorithms into a set of well-defined modules that can be flexibly assembled through a centralized controller. Under this scheme, hybrid schemes become an event-based search that can adaptively trigger a desired parents behavior in response to search events. MDF can hence be used to design and implement a wide spectrum of hybrids with varying degrees of collaboration, thereby offering algorithm designers quick turnaround in designing and testing their meta-heuristics. Such technicality is illustrated in the paper through the construction of hybrid schemes using ant colony optimization and tabu search.

Designing and tuning SLS through animation and graphics: an extended walk-through

Stochastic Local Search (SLS) is quite effective for a variety of Combinatorial (Optimization) Problems. However, the performance of SLS depends on several factors and getting it right is not trivial. In practice, SLS may have to be carefully designed and tuned to give good results. Often this is done in an ad-hoc fashion. One approach to this issue is to use a tuning algorithm for finding good parameter settings to a black-box SLS algorithm. Another approach is white-box which takes advantage of the human in the process. In this paper, we show how visualization using a generic visual tool can be effective for a white-box approach to get the right SLS behavior on the fitness landscape of the problem instances at hand. We illustrate this by means of an extended walk-through on the Quadratic Assignment Problem. At the same time, we present the human-centric tool which has been developed.