S. Muthuraman, C. MacLeod, and G. Maxwell (UK)
Artificial Neural Networks, Modular Networks, Evolution ary Algorithms, Robots, Locomotion
Artificial Neural Networks have so far failed to produce a convincing route to Robotic Intelligence. Training and Or ganizational Algorithms (such as Evolutionary Algorithms) are presently not flexible or sophisticated enough to config ure large networks which fuse data from different sensory domains in a complex and changing environment. The approach outlined here is different in that it al lows the neural network to grow, building itself up, piece by piece, from a simple to a complex form. This is accom plished by allowing the robot's body plan and environment to develop while simultaneously adding to the structure of the controlling network. Network structures from previous iterations are retained but are not retrained. Each time the robot attains a satisfactory performance with its current body plan in its current environment, com plexity is increased and new networks are configured on top of the old until this more challenging system is also mastered. The biological justification for this approach is out lined. Results are presented which demonstrate the opera tion of the approach in the development of a quadrupedal gait for a simulated robot.
Important Links:
Go Back
