Learning to Solve Bin Packing Problems with an Immune Inspired Hyper-heuristic

Motivated by the natural immune system’s ability to defend the body by generating and maintaining a repertoire of antibodies that collectively cover the potential pathogen space, we describe an artificial system that discovers and maintains a repertoire of heuristics that collectively provide methods for solving problems within a problem space. Using bin-packing as an example domain, the system continuously generates novel heuristics represented using a tree-structure. An novel affinity measure provides stimulation between heuristics that cooperate by solving problems in different parts of the space. Using a test suite comprising of 1370 problem instances, we show that the system self-organises to a minimal repertoire of heuristics that provide equivalent performance on the test set to state-of-the art methods in hyper-heuristics. Moreover, the system is shown to be highly responsive and adaptive: it rapidly incorporates new heuristics both when entirely new sets of problem instances are introduced or when the problems presented change gradually over time. Introduction Heuristic search methods have been shown to be successful in solving a wide-range of real-world problems. Typically for a given application domain, a range of heuristics for solving problems will exist; these might range in nature from simple rules encapsulating expert knowledge to complex search algorithms that need to be tuned by experts to work. Commonly, different heuristics will work well on problems in different parts of the problem space. By collecting together a set of heuristics, it is hoped that collectively, the weaknesses of individual heuristics can be compensated for by other heuristics in the set (Burke et al., 2003). The goal of the hyper-heuristics field is to find automated methods that can both generate appropriate sets of heuristics and provide a means of selecting between heuristics in the set, given either a new problem instance or even a partially solved problem instance. While such approaches have proved successful in many application areas, most hyper-heuristic approaches fail to continuously learn from experience On the one hand, the failure to exploit previous knowledge leads to inefficient hyper-heuristics; on the other, if the characteristics of instances of problems in the domain change over time, a hyper-heuristic may need to be completely re-tuned or in the worst case redesigned periodically. An ’ideal’ hyperheuristic would be able to exploit previous knowledge through access to some kind of memory, rapidly adapt existing knowledge to new circumstances, and additionally, generate new knowledge when previous knowledge is not applicable. We observe that the immune system fulfils very similar properties in its role as a host maintenance and defence system. The immune system maintains a repertoire of antibodies that has been shown theoretically to cover the space of potential pathogens. Clonal selection mechanisms provide a means of rapidly adapting existing antibodies to new variants of previous pathogens; meta-dynamic processes are able to generate novel antibodies; a memory mechanism enables the immune system to respond rapidly when faced with pathogens it has previously been exposed to. Using this analogy, we present a system that is shown experimentally to outperform single human-designed heuristics by a significant margin and furthermore, is shown to be more adaptable and responsive than a recent state-of-theart hyper-heuristic approach when faced with a continually changing problem landscape, thereby addressing the needs of real-world practitioners more fully. The novel system described has the following features: • it generates novel heuristics from a library of component parts • it utilises meta-dynamic processes to both add and remove heuristics from the system resulting in a self-organising network of heuristics • it sustains a network of interacting heuristics of minimal size that collectively solve problems from the whole of the problem space • it encapsulates memory in the sustained network enabling rapid adaptation to new problems Background We briefly cover some background in relation to the tested application domain of bin-packing and outline the immunolArtificial Immune Systems ICARIS