Sourodeep Bhattacharjee

Date: June 07, 2019

Time: 10 AM – 1 PM

Location: 3105, Lambton Tower

 

We have proposed an Estimation of Distribution Algorithm (EDA) based automated architecture search algorithm for combinatorial optimization named VAE-EDA-Q AVS. The proposed method uses Variational Autoencoders which allows uniform and smooth sampling of latent spaces and can efficiently explore the search space. VAE-EDA-Q AVS was tested on benchmark problems consisting of Trap Functions and NK Landscapes, and was found to have lowest execution time and fitness evaluations. VAE-EDA-Q AVS was extended to encode arbitrary Convolutional Neural Network (CNN) architectures of dynamically generated depths and to produce optimized offspring candidates of similarly varied CNN architectures. VAE-EDA-Q AVS was tested on CIFAR10 and CIFAR100 benchmark datasets successfully and compared to various state-of-the-art CNN algorithms. It was observed that VAE-EDA-Q AVS generates CNN models that have 1.5% higher accuracy for CIFAR10 on average compared to all other state-of-the-art algorithms while requiring 25% less parameters and 6% higher accuracy for CIFAR100 with 10% less parameters on average. 

Using individual based modeling platform known as EcoSim, we modeled the abilities of elitist mate selection and communication of fear. Data received from these experiments was reduced in dimension through use of VAE-EDA-Q AVS to act as a feature reducing wrapper method in conjunction with C4.5 Decision trees, which aided in further analyses of the data. Our results demonstrated a significantly lower speciation rate, a significantly lower extinction rate, for elitist mate selection group. Through modeling with EcoSim, we found alarm communication to decrease foraging activity in most cases, yet gradually increase foraging activity in some other cases.

 

Thesis Committee:

 

PhD Dissertation Announcement

 

5113 Lambton Tower, 401 Sunset Avenue, Windsor ON, N9B 3P4, (519) 253-3000 Ext. 3716, csgradinfo@uwindsor.ca