"ZOOM" SEMINAR

SCHOOL OF MECHANICAL ENGINEERING SEMINAR
Wednesday, March 24, 2021 at 14:00

Multi-Modal Multi-Objective Optimization with Solutions of Variable-Length

Yosef Breslav

M.Sc. student of Dr. Amiram Moshaiov

Abstract

Multi-modal optimization aims to provide decision-makers with alternative solutions, possibly near optimal, and not just one optimal solution. In recent years there has been a surge of interest on extending this approach to Pareto-based multi-objective optimization. This study appears to be the first one to focus on solving a special kind of Multi-modal Multi-objective Optimization Problems (MMOPs) in which solutions belong to decision-spaces of various dimensions (i.e., solutions of variable length).

First, a general problem definition for MMOPs with solutions of variable length is provided. Then, a solution framework is suggested to allow using standard Multi-modal Multi-objective Evolutionary Algorithms (MMEAs) to solve the considered type of problems. In addition, a novel MMEA is proposed. The algorithm, which is termed Archived Multi-Modal NSGA-II (amm-NSGA-II), constitutes a modification to the well-known NSGA-II to allow solving MMOPs.

Following the algorithm description, a real-life example with solutions of variable-length is suggested concerning optimal antennas’ layout-allocation design for a wireless communication network. Next, the proposed amm-NSGA-II is employed to solve three different optimal antennas’ layout-allocation design problems. When compared with other MMEAs, it is shown that the proposed algorithm provides not only better solution diversity in the decision-spaces, but also solutions with superior performance vectors. This superiority appears to be attributed to the type of archive that is used here, which leads to recommendations concerning future research.  

Join Zoom Meeting

https://us02web.zoom.us/j/88626960530?pwd=K0NRR3dheVcxRVUvVWtyRHhNTFZXQT09

Note: The seminar will be given in Hebrew.

Estimating Information-Theoretic Measures with Classification Algorithms

Yuval Shalev

  Ph.D candidate at the Department of Industrial Engineering in the Tel-Aviv University, 

  under the guidance of Prof. Irad Ben-Gal and Dr. Amichai Painsky.
 
https://us02web.zoom.us/j/88556814331?pwd=REJvREpGNzVJUlQ4Y0lRSFczTHJkUT09
 
   

Abstract:
 
Estimating the information-theoretic measures of discrete random variables is a fundamental problem in information theory and related fields. This problem has many applications in various domains, including machine learning, statistics and data compression. Over the years, a variety of estimation schemes have been suggested. However, despite significant progress, most methods still struggle when alphabet size is large, random variables are from high dimensional space or when time dimension is involved. In this work, we introduce a holistic solution for this problem that is based on classifiers models. A set of these models is trained on the joint distribution of the random variables ,to obtain estimators for fundamental measures such as entropy, mutual information and transfer entropy. We focus on estimating these measures in scenarios where other methods usually fail. We  provide an experimental study of several use cases that demonstrates the advantages of our proposed scheme over state-of-the-art methods.
 
Bio:
 
Yuval Shalev is a Ph.D candidate at the Department of Industrial Engineering in the Tel-Aviv University,   under the guidance of Prof. Irad Ben-Gal and Dr. Amichai Painsky. Previously, he completed his M.Sc studies in physics in the Hebrew University. His research interests include Machine learning, deep learning and their connection to information theory. In parallel to its academic studies, Yuval has been working in the financial industry since 2008 as a quant and a data scientist.  Today, He leads the research team in the data science group of Citi innovation lab in Israel.