| |
Exploring the T-Maze: Evolving Learning-Like Robot Behaviors using CTRNNs
|
|
5 Experiment 3: Double T-Maze
In the previous section it was shown that evolved robots were able to completely
solve the simple T-Maze even in the case when the reward position was switched
during an epoch. However, the robots only had to retain one piece of information
based on previous experiences, namely whether to turn left or right at the
T-junction. The investigations were now turned towards a double T-maze with
several T-junctions thus further complicating the task (see figure 8). To compensate
for the increased size of the maze the number of sensory-motor cycles
was increased to 400 in trial 1 and 200 in the remaining trials. The reward-zone
could appear in either upper-left and upper-right corner of the maze.
During the first trial only the reward-zone was present, and in the following
trials poison-zones were placed in the 3 remaining corners. The other parameters
remained unchanged. As in the simple T-Maze, the case where the reward
position remained unchanged during an entire epoch was tested first. An evolutionary
process seeded with populations of random individuals was launched,
but the results were very poor under these conditions. Basically the fitness remained
at zero all the time, with very few exceptions of fitness 1 coming and
going for a couple of generations in one of the replications of the experiment.
In fact this result is not that surprising considering the fitness function used.
In order for a random initial individual to collect some fitness and kick-off the
evolutionary progress it had to navigate all the way to one of the upper corners.
If it could not do that it simply got zero fitness. A solution to this problem
could have been to design an incremental fitness function, where individuals
would get some fitness for partially solving the task. Instead it was decided to
again rely upon an incremental evolutionary approach. The genetic algorithm
was seeded with a population consisting of the best individuals from one of the
runs of experiment 1 on the simple T-Maze. The results of 10 replications of the
experiment are shown in figure 9. In the best run the fitness reached a level of
18 out of 20. During trial 1 the best individual always turned right at the first
junction and left at the second. This would take the robot to the upper right
corner of the maze. In epochs with reward to the right the robot would soon
find and stay in the reward-zone. When the reward was on the left, on the other
hand, the robot was not capable of searching for the reward-zone as it did in
the simple T-maze case (fig. 4(a)). Instead the robot simply crashed into the
upper wall. In the following trials of these epochs, however, the robot would now
turn left at the first junction and right at the second, reaching the reward-zone
in the upper-left corner. One trial was wasted, but crucial information about
the reward position was gathered and used in the following trials, thus resulting
in fitness of 18 out of 20. An attempt to further improve this behavior by an
additional incremental evolution was now conducted, seeding evolution with a
population of individuals from this experiment. This time, however, the attempt
was not successful and the fitness level remained at 18 in every replication of the
experiment. This results suggests that the maximal problem complexity solvable
for the genetic algorithm and neural network used had been reached. This
suggestion was confirmed by an unsuccessful attempt to apply reward-switching
to the double T-Maze task, as was done in experiment 2 in the simple T-Maze
case. No reliable learning behaviors were observed in the evolved controllers in
this case.
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
|
|
|
|
