|
The experiments are carried out using a standard genetic algorithm with
rank-based selection. A population of 200 randomly generated neural controllers
is evolved for 200 generations. At every generation the best 40 individuals make 5
copies each. One copy each of the 5 best individuals remains unchanged (elitism).
For the rest of the population single-point crossover with a probability of 0.04
and bit-switch mutation with a 0.02 probability per bit is applied. The whole
experiment is repeated 10 times using different initializations of the computer’s
pseudo-random number generator.
The fitness results of the evolutionary runs on this experiment are shown in
figure 3. The thick line shows best fitness and the thin line shows population
mean, both are averaged over 10 replications of the experiment. The evolutionary
process found individuals able to collect the maximal fitness of 20 in 6 out of
the 10 replications of the experiment. The maximal fitness in the 4 remaining
runs was around 16. The behavior of an individual from the final generation of
one of the successful runs is shown in figure 4 and 5. The robot starts out in
trial 1 of the first epoch (figure 4(a)) by exploring the maze until it locates the
reward-zone where it stays the remaining time of the trial. In the following trials
(figure 4(b)), the robot is able to retain the “knowledge” gathered during trial
1 and always turns left at the T-junction in order to move towards and stay on
the reward-zone. In the epochs with reward to the right the robot moves directly
towards the reward-zone in trial 1 (figure 5(a)), since the default behavior of this
individual is to turn right at the first junction after an re-initialization of the
neural controller. For the remaining trials this successful behavior is repeated
(figure 5(b)).
3.1 Analysis
In order to better understand the functioning of the evolved neural controllers,
some further analysis on an individual from one of the successful runs was done.
The neural activities of each neuron were recorded over two epochs - one with
reward to the left and one with reward to the right. It was found that the
essential information about the current environment is stored in one of the hidden
neurons. The activity of this neuron approaches zero at the end of the trials with
the reward to the left and approaches one otherwise. By initializing every other
neuron in the network as normal, but setting this neurons activity to either
zero or one, it could be controlled which way the robot turns at the T-junction.
In other words the state of this neuron “stores” the robots current assumption
about the environment and is updated when these assumptions are not met. For
more details about the analysis performed please refer to [1].
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
|
