Multi-Objective Deep Network-Based Estimation of Distribution Algorithm for Music Composition

Abstract

In the field of evolutionary algorithm music composition, most of the current researches focus on how to enhance environmental selection based on multi-objective evolutionary algorithms (MOEAs). However, the real music composition process defined as large-scale multi-optimization problems (LSMOP) involve the number of combinations, and the existing MOEA-based optimization process can be challenging to effectively explore the search space. To address this issue, we propose a new Multi-Objective Generative Deep network-based Estimation of Distribution Algorithm (MODEDA) based on dimensionality reduction in decision space. In order to alleviate the difficulties with dimensional transformation, we propose a novel solution search method that optimizes in the transformed space and ensures consistency between the pareto sets of the original problem. The proposed algorithm is tested on the knapsack problems and music composition experiments. The experimental results have demonstrated that the proposed algorithm has excellency in terms of its optimization performance and computational efficiency in LSMOP.

Estimation of Distribution Algorithms in Machine Learning: A Survey

“…In most EDAs that do not restrict the dependence relationships between variables, the joint probability distribution is estimated by a Bayesian network (Section II-C) learned from data. EDAs have also been developed with the probability distribution estimated from log-linear probability models [39], probabilistic principal component analysis [40], Kikuchi approximations [41], Markov networks [42], [43], Markov chains [44], copulas and vines [45], a reinforcement learningbased method [46], Gaussian adaptive resonance theory neural networks [47], growing neural gas networks [48], restricted Boltzmann machines [49], [50], [51] and in the deep learning area, from autoencoders [52], variational autoencoders [53], [54], and generative adversarial networks [55]. Model selection in EDAs is a more complex problem.…”

“…Most of these algorithms, EDAs included, simplify the problem by reducing the m-dimensional space to a scalar value with fitness functions like the convergence indicator, the Pareto-optimal front coverage indicator, the hypervolume indicator and the unary additive -indicator. This is the strategy followed by EDAs based on neural networks [47], [48], [54], [51], on probabilistic models [82], [103], [107] or on a Parzen estimator [108].…”