school of mechanical engineering Gil Segal and Adham Salih

13 בדצמבר 2017, 14:00 - 15:00 
 
ללא תשלום
 school of mechanical engineering Gil Segal and Adham Salih

 

 

 

 

School of Mechanical Engineering Seminar
Wednesday, December 13, 2017 at 14:00
Wolfson Building of Mechanical Engineering, Room 206

 

 

Comparing Algorithms for Multi-objective Evolution of Neuro-Controllers

 

Adham Salih

MSc Student of Prof. Ami Moshayov

 

In recent years there has been an increase of interest in designing Neuro-Controllers (NCs) using multi-objective evolutionary computation techniques. Given the vast variety of Multi-Objective Evolutionary Algorithms (MOEAs), selecting one for a specific problem is a non-trivial task. The presented study aims to provide a comprehensive comparison between two well-known MOEAs including NSGA-II and MO-CMA-ES. Given the fundamental difference between the selection and reproduction mechanisms of these two algorithms, it should be asked which of these algorithms is better for the multi-objective evolution of NCs. Past studies on neuro-evolution point at a possible convergence difficulty, when crossover mechanism is used to create offspring. This difficulty, which is attributed to the well-known competing convention problem, may lead to a conclusion that MO-CMA-ES should be preferred as it is based on mutation rather than crossover.

First, a methodology is suggested for testing and comparing MOEAs as applied for multi-objective evolution of NCs. Next, based on the suggested methodology, an extensive comparison is carried out between the two tested algorithms. The investigation includes five benchmark problems, each with different control difficulties. Statistical analysis of the results indicates that none of the tested algorithms is superior with respect to all the considered problems. Possible explanations to the observed results are discussed and suggestions for future research are provided.

 

Learning the behavior of an individual locust in a swarm by an Adaptive Neuro Fuzzy Inference System (ANFIS)

Gil Segal

MSc student of Dr. Amiram Moshaiov and Prof. Amir Ayali

 

A locust swarm is an exceptional example of a coordinated motion in nature. The general motivation for this study is the desire to understand how the behavior of an individual in such a swarm is translated into a collective movement. This study deals with the first step towards answering this question. It concerns identification of the behavior of an individual locust in a group of marching locusts, as observed in laboratory conditions. The identification is performed using the Adaptive Neuro-Fuzzy Inference System (ANFIS) method. In contrast to the use of artificial neural networks, employing ANFIS allows building a system that is interpretable and easy to analyze. This advantage of ANFIS is useful for understanding the complex biological system that is dealt with here.

Clearly, some major uncertainties are inherent to our understanding of the aforementioned system, as is the case with any other biological system. A major concern is the lack of knowledge about which are the inputs that a locust takes into account, when a particular motion action is experienced. Observing locusts in a swarm, it can easily be noticed that the motion is intermittent and an individual often stops for certain periods of different time-spans. We focus on the individual’s decision to stop or resume walking, i.e. to join the collective movement. The main assumption is that if a reasonable ANFIS-based model is found for an individual locust, then it may hint at the inputs that are actually used by the locust, as related to the behavior of the swarm. Based largely on some knowledge of the biological sensory system of the locust, a trial-and-error approach has been used. It resulted with a bio-plausible set of inputs that provided substantially better identification results, as compared with several other such sets. We achieved a controller that succeeded in predicting the behavior of the individual locust with a certainty of more than 96%. The controller that succeeded in reaching this certainty had a reaction time of two frames (approximately 70 msec), and was based on three inputs: distance to the nearest obstacle, locust velocity and the number of walking neighbors behind the locust.

Future work may include learning the velocity vector of the locust. This may lead the way to the prediction of the entire motion of an individual locust, and could serve as a base for a simulation of the coordinated motion of the locust.

 

 

 

 

אוניברסיטת תל אביב עושה כל מאמץ לכבד זכויות יוצרים. אם בבעלותך זכויות יוצרים בתכנים שנמצאים פה ו/או השימוש
שנעשה בתכנים אלה לדעתך מפר זכויות, נא לפנות בהקדם לכתובת שכאן >>