Estimation of Distribution Algorithms (EDAs) have been successfully applied to a wide variety of problems. The algorithmic model of EDA is generic and can virtually be used with any distribution model, ranging from the mere Bernoulli distribution to the sophisticated Bayesian network. The Hidden Markov Model (HMM) is a well-known graphical model useful for modelling populations of variable-length sequences of discrete values. Surprisingly, HMMs have not yet been used as distribution estimators for an EDA, even though it is a very powerful tool especially designed for modelling sequences. We thus propose a new method, called HMM-EDA, implementing this idea. Preliminary comparative results on two classical combinatorial optimization problems show that HMM-EDA is indeed a promising approach for problems that have sequential representations.
Categories and Subject Descriptors
KeywordsEstimation of distribution algorithms; Hidden Markov models; Combinatorial optimization Several graphical models have been used in EDAs. The most widely known is probably the Bayesian optimization algorithm (BOA) [3], which relies on a Bayesian network to learn the model and sample solutions, but there are more, like DEUM (Distribution Estimation Using Markov network) [5] and dtEDA (dependency-tree EDA) [4]. HMM-EDA is a novel approach which uses a Hidden Markov Model (HMM) [2] as the underlying distribution. While its versatility may allow it to model various kinds of distributions, the focus of this work is on combinatorial optimization problems.The proposed approach integrates an HMM in the generic EDA model, using it to directly estimate the distribution of the samples making up the population. Using an HMM for that purpose is quite versatile, as it only assumes that the samples are sequences of discrete values. Such a model should be useful in a wide variety of optimization problems, including problems where solutions can be modelled as bit strings. HMM should be able to capture complex interactions between the elements of sequences, not only the first order relations captured in observable Markov models. Our intuition is that hidden states of the HMM can capture intrinsic sequential patterns that characterize populations, and link these patterns together in various ways to produce better individuals, through transitions between the hidden states.The HMM-EDA algorithm interprets individuals as sequences. Each possible value of the problem's "alphabet" is mapped to a possible observation of the HMM. For instance, with the travelling salesman problem, there would be as many possible observations as the number of cities in the tour. Given that HMMs can easily handle up to several dozens of observable emissions, the implementation is quite straightforward for many problems. The HMM is trained with the best individuals, and then produces a new generation, effectively leading to an offspring population with, hopefully, a better average fitness than the parent population. In the process, the probabilistic nature of the HMM will a...