Research Preparation Exam: Reza Mahjourian, GDC 3.816

Contact Name: 
Lydia Griffith
May 15, 2014 1:00pm - 3:00pm

Research Preparation Exam: Reza Mahjourian

Date: Thu May 15
Time: 1pm
Place: GDC 3.816
Committee Members: Risto Miikkulainen, Al Mok, Luis Sentis

Title: Robotic Control through Neuroevolution


Traditional industrial robotics is primarily focused on optimal path
control. In contrast, modern robotic tasks increasingly require greater
degrees of adaptability and flexibility to deal with variations in the
robot's environment.  Moreover, new robotic tasks can be specified
without precise models or target behaviors.  Neuroevolution of robotic
controllers is a promising approach to such tasks: Neural networks
represent nonlinear control naturally and evolutionary computation
creates robust and flexible solutions.  In this talk I will present a
brief review of existing work on neuroevolution of robotic controllers.
I will then discuss early experiments using a neuroevolution method
called NEAT to evolve a controller for an object manipulation task.
Results show that the choice of neural network architecture and task
specification make a difference. I suggest possible improvements on them
in order to scale up the approach towards flexible partially autonomous
humanoid control in the future.