Our mission is to explore and establish a methodology of evolutionary intelligence that produces a highly adaptive, autonomous, and theoretically reliable learning and optimization ability. The current research topics include the followings:
- Expensive optimization
Many real-world applications require optimizing expensive-to-evaluate objectives, as objective values are evaluated often with computationally expensive simulations or costly experiments. An example is the optimization of aircraft wing using CFD simulation, which is computationally expensive. To solve such expensive optimization problems, we explore sample-efficient optimization techniques. Especially, high-dimensional multi-objective problems are our main focus to be studied.
- Evolutionary machine learning
Integrating meta-heuristics into machine learning approaches bring highly adaptive and autonomous ability, such as auto-design of learning models, auto-tuning of hyper-parameters, and extraction of explainable models. We explore evolutionary machine leaning techniques; evolutionary rule-based learning, evolutionary symbolic regression, evolutionary neural architecture search, and etc. We are interested in realizing the interplay between evolution and learning on a computer intelligence scheme.
05 April 2023
We have proposed a novel approximation-based multiobjective evolutionary algorithm to solve high-dimensional expensive multi-objective optimization problems. A main finding is that constructing an approximation model for each single-objective decomposed problem can be a reasonable strategy to address the high-dimensionality of problems. The proposed algorithm can drive state-of-the-art performance compared to modern algorithms. This contribution has been released on TechRxiv as a preprint document. paper code
14 March 2022
multiobjective evolutionary algorithm, which utilizes classification models adapted to single-objective decomposed problems, to solve high-dimensional expensive multi-objective optimization problems. Experimental results confirm that the proposed algorithm is competitive to the state-of-the-art algorithms and computationally efficient as well. This contribution has been published on IEEE Transactions on Evolutionary Computation. paper code
13 February 2021
Evolutionary Rule-based Learning (ERML) approaches have been applied to modern issues in the machine learning field. We have derived a theoretical approach that mathematically guarantees that ERML identifies maximally accurate rules in the fewest iterations possible, which also returns a theoretically valid hyper-parameter setting. We also experimentally show that our theoretical setting enables ERML to easily solve several challenging problems where it had previously struggled. This contribution has been published on IEEE Transactions on Evolutionary Computation. paper code
Our laboratory is on the eighth floor of N6-2, Hodogaya-campus, Yokohama National University. See here for more detail.
Faculty of Engineering, Yokohama National University
Room 801/812, N6-2, Tokiwadai 79-5, Yokohama, Japan, 240-8501.
nakata-masaya-tb at ynu.ac.jp（to Masaya Nakata）