Results
From MarsRoverSim
Contents |
[edit] Preliminary Selection & Mutation Tests
Initial test of the neural network and PaGMO to establish the most appropriate selection/mutation combination. This image shows different configurations of GA parameters and their influence on the performance of the GA. In this particular case the GA was maximising the output of the neural network.
[edit] Island & Migration Comparison
This graph shows the maximum (blue) and maximum mean values (red) for 4 repetitions of 8 islands with 10 individuals, compared with the maximum values obtained with the same settings with no migration (purple) and one island of 80 individuals (green). All other starting parameters remained the same. (i.e. the values plotted are the best maximum and best average for each generation).
This graph is similar to the graph above, the only difference is that all the values are averaged from all the four runs (e.g. any max fitness data point is the average of max fitnesses from 4 runs).
Blue lines = 1 island with 80 individuals; Red lines = 8 islands with 10 individuals each; orange lines = no migration, 8 islands with 10 individuals each. Dashed lines show the average and full lines show the maximum fitness.
This graph is a comparison of the effects of migration in eight islands. Here we have the data from 10 separate evolutionary runs. Red lines show runs with migration enabled and blue lines shown those without migration.
This graph compares 8 Islands (8x10 individuals) with and without migration to 1 Island (1x80 individuals).
This graph with error bars represents the same data as the above graph.
This graph shows the best and average fitness for 1 island, 8 islands with migration, and 8 islands without migration. Results averaged from 20 replications.
[edit] Migration Experiments
[edit] 17/11/09
[edit] 18/11/09
[edit] 19/11/09
[edit] 20/11/09 (One Island)
[edit] 21/11/09 (No Migration)
[edit] 27/11/09 - 04/12/09
[edit] Multiple Terrains
[edit] 28/01/10 (Multiple Terrains I)
[edit] 01/02/10 (Multiple Terrains II)
[edit] 05/02/10 (Multiple Terrains III)
[edit] Comparison
Controllers from the above multiple terrain experiments were compared with controllers from previous experiments (evolved in a single terrain) to see whether the settings used in the multiple terrain methodology produced greater generality (i.e. ability to cope with new environments).
Controllers were analysed over 10 terrains for 50 trials - the averages can be seen in the graph below, clearly showing that the multiple terrain evolution reduced controller generality.
[edit] Active Vision 1
[edit] 15/02/10 - Active Vision 1 (Low Mutation)
[edit] Island Size Comparison
An experiment was performed to see the effects of islands on populations of different sizes, with three different conditions: 8 islands with migration, 8 islands without migration, and single populations.
80 individuals:
160 individuals:
240 individuals:
Comparison:
[edit] Analogue Sensors
Controllers were evolved using continuous, floating-point sensors in a simple environment with 2 rocks and 2 holes.
The below graph shows the average and average maximum evolution fitness over 10 replications.
Individuals at the final generation were then evaluated for 100 trails in the evolution terrain:
As well as two new environments:
[edit] Active Vision 2
Using an environment with different terrain properties and an illuminated landmark, an experiment was conducted to investigate island evolution to evolve visual landmark and obstacle avoidance behaviours.
Graph of evolution for 5 replications:
Example trajectory, showing visual navigation:
