Download presentation
Presentation is loading. Please wait.
Published byMadeline Ferguson Modified over 8 years ago
1
Evolving robot brains using vision Lisa Meeden Computer Science Department Swarthmore College
2
Can be used in a variety of classes Introduction to Cognitive Science Students observe and describe results of evolution Artificial Intelligence Students modify the evolutionary process and report on the different outcomes Robotics Students use this example as a base for designing their own projects
3
Genetic Algorithm Start with a random population of individuals For each generation of the evolution process: Fitness proportionate selection Reproduction Mutation Repeat until best member of population is good enough
4
Framework Pyrobot simulator Green robot Evolving brain Sensors: camera and sonar Red robot Fixed brain: move straight and avoid obstacles Sensors: sonar
5
Genetic Algorithm Details Evolve the weights of a fixed size 3-layer neural network that maps sensors to motors Initialize 10 neural networks with random weights Allow robot to move for 250 steps, fitness based on: Absolute value of translation speed Whether the robot is stalled Centeredness of red blob in camera image Closeness of red blob in camera image Evolve for 10 generations, saving the best weights from each generation
8
Conclusions Using vision in a very simplified way, but it enables students to appreciate the power of evolution in a relatively short demonstration Evolving neural network weights, rather than using fully supervised algorithms such as back- propagation, allows students to create more open- ended robot learning problems
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.