All faculty and students are cordially invited to attend the Doctoral Dissertation Defense for Ahmed Aly.
Zongli Lin, ECE
Gianluca Guadagni, ES
Mike Gorman, ES
Gabe Robins, CS
Advisor: Joanne Bechta Dugan, ECE
Title: Experimental Studies in Pursuit of Experiential Robot Learning
The topic of Robot Learning has been investigated and based upon our observations a method to train robots called Experiential Robot Learning (ERL) has been proposed. The ERL method is poised to address a gap we identified in the literature. The problem is that robots are not mature enough to be used in unconstrained environments (i.e. in the wild) because they cannot learn and thus cannot respond to new situations. Our hypothesis therefore is that the development of a methodology that permits experiential learning could allow robots to learn and therefore to succeed in novel situations.
What are the requirements on experiential robot learning that would enable robots to succeed in novel situations? ERL would need to be experiential, open ended, scalable and platform agnostic, among many other characteristics. In this regard, Neural Networks (NN) provide a promising path towards ERL and this dissertation evaluates this promise.
The early experiments illuminated a problem with using Deep Learning for ERL: the need for differentiable, informative loss functions cannot always be satisfied. The exploitative behavior of gradient-following can also be incompatible with the exploratory nature of ERL.
To address these shortcomings, we developed 3 gradient-free algorithms called Accelerated Neuroevolution (AN), Multiple Search Neuroevolution (MSN) and Local Search (LS). All three algorithms were developed, and validated, to train deep neural networks and provide good results on the experiments performed. The experimental studies had a wide variety including classic control tasks, solving Global Optimization functions, Image Classification on FashionMNIST dataset and physical tasks with the NAO Humanoid robot. AN successfully trained a robot to rotate its head to face objects. MSN solved MNIST using a 5 Million-parameter CNN. LS solved all Classic Control tasks in Open AI Gym.
In conclusion, we paved way to addressing shortcomings of hand-coding approaches with different arcs of contributions. First is that we developed a high-level method to outline how and in what way robots should be developed. Secondly, we conducted experimental studies to address certain aspects in that framework, such as investigating new neural network training techniques or producing unique hybrid neural architectures. The experimental studies generated new insights on how to train neural networks following the ERL method. A by-product contribution came as a major software framework to train NNs on very different tasks by only changing a handful of lines of code.