Date of Award

Summer 8-15-2017

Author's School

Graduate School of Arts and Sciences

Author's Department

Physics

Degree Name

Doctor of Philosophy (PhD)

Degree Type

Dissertation

Abstract

We study machine learning class classification problems and combinatorial optimization problems using physics inspired replica approaches. In the current work, we focus on the traveling salesman problem which is one of the most famous problems in the entire field of combinatorial optimization. Our approach is specifically motivated by the desire to avoid trapping in metastable local minima-a common occurrence in hard problems with multiple extrema. Our method involves (i) coupling otherwise independent simulations of a system (“replicas”) via geometrical distances as well as (ii) probabilistic inference applied to the solutions found by individual replicas. In particular, we apply our method to the well-known “k-opt” algorithm and examine two particular cases-k = 2 and k = 3. With the aid of geometrical coupling alone, we are able to determine for the optimum tour length on systems up to 280 cities (an order of magnitude larger than the largest systems typically solved by the bare k = 3 opt). The probabilistic replica-based inference approach improves k - opt even further and determines the optimal solution of a problem with 318 cities. In this work, we also formulate a supervised machine learning algorithm for classification problems which is called “Stochastic Replica Voting Machine” (SRVM). The method is based on the representations of known data via multiple linear expansions in terms of various stochastic functions. The algorithm is developed, implemented and applied to a binary and a 3-class classification problems in material science. Here, we employ SRVM to predict candidate compounds capable of forming cubic Perovskite structure and further classify binary (AB) solids. We demonstrated that our SRVM method exceeds the well-known Support Vector Machine (SVM) in terms of accuracy when predicting the cubic Perovskite structure. The algorithm has also been tested on 8 diverse training data sets of various types and feature space dimensions from UCI machine learning repository. It has been shown to consistently match or exceed the accuracy of existing algorithms, while simultaneously avoiding many of their pitfalls.

Language

English (en)

Chair and Committee

Zohar Nussinov

Committee Members

Anders Carlsson, Li Yang, Ralf Wessel, Joseph O'Sullivan,

Comments

Permanent URL: https://doi.org/10.7936/K77080VN

Download

Included in

Physics Commons

Share

COinS