סמינר מחלקה של דוקטורנט אדהם סליח - אבולוציה מרובת מטרות של טופולוגיה ומשקלות של רשתות עצביות
School of Mechanical Engineering Seminar
Wednesday, December 22, 2021, at 14:00
Wolfson Building of Mechanical Engineering, Room 206
Ph.D. student of Dr. Amiram Moshaiov
Many-objective Topology and Weight Evolution of Artificial Neural Networks
Neuro-Evolution (NE) combines the adaptation power of Artificial Neural Networks (ANNs) with the advantages of evolutionary computation techniques to find networks capable of solving different tasks. The potential of NE has been successfully demonstrated in many studies involving application areas such as robotics, artificial life, computer games, agent technologies, and computer vision. A major advantage of the evolutionary approach to designing ANNs is its ability to simultaneously search for both the optimal topology and weights of the ANNs. This type of NE is known as TWEANN (Topology and Weight Evolution of Artificial Neural-Networks). Over more than a decade, the significance of using multi-objective optimization for the design of ANNs has become increasingly apparent from various studies. It appears that most TWEANN studies are restricted to either single objective problems or problems involving up to three objectives.
This study aims to develop and investigate many-objective TWEANN algorithms, i.e., algorithms for solving problems with more than three objectives. First, a general decomposition framework approach has been suggested for the development of many-objective TWEANN algorithms. The attractiveness of using a decomposition approach is that it is a leading approach for dealing with many-objective optimization problems. This approach allows not only the increase in the problem dimensionality but also to bias the search according to the objective preferences of the user. Next, the proposed framework has been employed to devise a specific TWEANN algorithm, which has been termed as Neuro Evolution of Weights and Structures by Decomposition (NEWS/D). Furthermore, NEWS/D has been modified into a neuro-fuzzy version, which has been termed as Fuzzy Evolution of Membership and Structure by Decomposition (FEMS/D).
To evaluate the capabilities of NEWS/D to handle many-objectives, a systematic approach has been suggested to validate many-objective TWEANN algorithms. This approach involves various types of benchmark problems including problems with discrete and continuous outputs. To demonstrate the usefulness of NEWS/D and FEMS/D, several many-objective robot motion-control problems have been devised. These include: 1. A meta optimization problem to find controllers that have satisficing performances in up to eight different environments; 2. A many-objective problem to promote the transferability of controllers to unseen environments; 3. An adaptation of neuro-fuzzy controllers to an abrupt change in the environment; and 4. A dynamic three-dimensional pursuit-evasion aerial problem. Finally, in this presentation, several results out of these four different problems will be provided to illustrate the applicability of both NEWS/D and FEMS/D.