Machine Learning

Machine learning methods are developed, including methods for the analysis of big data. In particular, methods for classification of users of online resources are proposed.

Methods for solving machine learning problems are being developed using SAT-solvers and CSP-solvers, as well as swarm intelligence techniques. Theoretical study of evolutionary algorithms is being performed, including time analysis of evolutionary algorithms adjusted with reinforcement learning.

ADAPTIVE SELECTION OF EXTRA OBJECTIVES USING REINFORCEMENT LEARNING

Single-objective optimization can be enhanced by adding auxiliary objectives, but how should we choose the most efficient ones, and when should we use the particular objective? A method designed to solve these issues was proposed in our laboratory [1-3]. The method is called EA+RL, which stands for an evolutionary algorithm (EA) controlled with reinforcement learning (RL).
 
There are several techniques that involve using some additional objectives in order to enhance performance of EAs. In multiobjectivization technique [6] all the objectives are optimized simultaneously by some multiobjective algorithms (MOEAs). In this technique the objectives should be specially developed in order to increase the optimization performance. It was shown that adding an inefficient objective leads MOEAs to fail on the considered model problems [3].
 
Helper-objective approach also involves using MOEAs, but it requires a strategy of choosing the auxiliary objective to be optimized at the current population [5]. The strategy can be either random, or ad-hoc [7]. The random one is general, but it does not take advantage of problem characteristics. At the same time, ad-hoc strategies can be efficient, but they lack generality. 
 
The EA+RL method incorporates auxiliary objectives into a single-objective EA. It requires less computational effort than MOEA-based methods, which makes it more applicable to resource-consuming problems. The selection strategy used in EA+RL is problem independent and it allows to learn some features of the problem as well, thus the method seems to increase both efficiency and generality of the helper-objective approach. 

Information © 2015-2017 ITMO University
© 2015 Department of Information Technology