Tensor Network Estimation of Distribution Algorithms

There is a long history of interaction between machine learning models and evolutionary algorithms going at least as far back as the 1970s [1]. This interaction has gone in both directions [2]. Evolutionary algorithms have been used to aid machine learning, for example, to tune hyperparameters of convolutional neural networks [3], search over model architectures [4], or optimize parameters of neural networks in reinforcement learning algorithms [5]. And conversely, machine learning models have been used to aid evolutionary algorithms, often by forming components of a larger evolutionary algorithm. Examples abound: machine learning models have been used to define fitness functions [6], to define chromosome representations [7], to perform "smart" mutations [8], and to generate new individuals from old by, in essence, performing a sophisticated form of crossover. An early example of machine learning models used as a quasi-crossover component are Estimation of Distribution Algorithms (EDAs) [9, 10, 11, 12, 13]. EDAs are a family of optimization algorithms where "parent" solutions are used to fit or update a generative model from which "children" solutions are sampled. Better solutions are then selected from among the children to become the next "parents", i.e. the training data for the generative model of the next iteration.