|
6 Conclusion
We have shown that evolution of learning-like properties is possible without
modifications of synapse strengths, but simply by relying on complex internal
dynamics of CTRNNs. In experiment 1 the robot had to navigate a simple TMaze
with the reward position fixed during each epoch. Direct evolution of this
task was possible and the analysis showed that the employed strategy of an
evolved network was to store essential environment information in one of the
hidden units. The rest of the neurons would update the activity of this neuron
based on current environmental feedback. With a few simulator modifications
evolved behaviors were successfully transfered to a real robot. In experiments 2
the reward position would vary within the same epoch forcing evolved robots to
keep on adapting their strategy to the current environmental conditions. Successful
individuals solving this task were evolved in a two-step incremental evolution.
Because of the increased complexity of the maze, direct evolution was not
possible in experiment 3. However, by seeding evolution with a population from
experiment 1 individuals capable of “learning” were found. It was not possible to
perform additional reward switching as in experiment 2 on the simple T-Maze,
suggesting that maximal task complexity had been reached.
In this work incremental evolution has proven to be a powerful tool in evolving
complex Robot Behaviors however evolving CTRNNs as shown here will face
evolvability problems if the task complexity is to be further increased. In principle
a sufficiently large CTRNN is able to display arbitrarily complex dynamics.
However, the problem of how to evolve such networks will have to be addressed
in the future. Possible solutions could be to explore new neural mechanisms for
information “storage”, or to investigate how to preserve the learning capabilities
of CTRNNs in networks generally thought of to be easier to evolve such as
Hebbian synapse networks or spiking neural networks. Our work will focus on
these aspects in the future.
As pointed out by one of the reviewers it can be argued whether the experiments
presented in this paper should be classified as learning-like behaviors or
simply as internal dynamics investigations. It true that the view of learning presented
in this paper is quite different from the traditional computational view
of learning, where some update of the Control Systems always takes place. However
neurophysiological experiments have indicated that the way animals and
humans perceive, classify, and memorize, for example in the olfactory system, is
by transitions between chaotic attractors in dynamical systems formed by large
numbers of neurons in the brain [3][7]. These results correspond nicely with the
view of memory and learning presented in this paper.
Page 1, 2, 3, 4, 5, 6
Tech Materials (Free)
Robot Behaviors Exploring the T-Maze: Evolving Learning-Like Robot Behaviors using CTRNNs
Humanoid Robotics A Biochemical Subsystem for a Humanoid Robot
Industrial Automation Systems Applying Agents for Engineering of Industrial Automation Systems
Robot Team Cooperation A Descriptive Model of Robot Team and the Dynamic Evolution of Robot Team Cooperation
Kuka Robots For ONU ONU Robotics Technology Center of Excellence, powered by KUKA Robotics Corporation
Augmented reality Annotation System for Robotic Application
Modular Robots Self-Reconfiguration Planning Of Identical Modules
Autonomous robots A New Approach To Robotics
Robotic Mounting Flat Panel Displays With Robotic Mounting
Calibration of Industrial Robots A Photogrammetric Robot Calibration System Based On Off-The-Shelf Low Cost Hardware Components
More... Amazon Books
|
