Teams of Coevolved Predators in a Predator/Prey Grid World
This page presents research in a predator/prey torus grid world done by undergraduate student Alex Rollins as
part of Southwestern University's Summer research program
SCOPE.
The research makes use of the MM-NEAT software package
(an extension of NEAT),
which has been extended to allow for coevolution of distinct subpopulations. Predator agents are controlled
by evolved neural networks.
Specifically, these experiments evaluate the relative merits of having team-based objectives as opposed to individual-based objectives (or both),
and of having a single mode of behavior available as opposed to having two (as a result of separate neural output modules).
Each video shows three complete trials. The predators included in each video are the champion team of generation 300 for the corresponding experimental setup.
The videos on this page can be viewed in a playlist here.
Team Fitness Only: 1 Module
This video is of the Team1M predator team, where the predators are given only team-based fitness functions and have unimodal neural networks. The color trails seen behind each predator track where the predators have been most recently. This team did not develop specializations very strongly, as all three predators are almost always moving aggressively towards the prey and none of the predators are taking on a support role (such as a blocker). Still, on the trials that are shown in this video, the predators accomplished the following by the end of each trial:
This video is of the Individual1M predator team, where the predators are given only individual-based fitness functions and have unimodal neural networks. The color trails seen behind each predator track where the predators have been most recently. This team was able to develop specializations much better than the previous team (Team1M). This is demonstrated in the video with 2 predators acting as aggressors and 1 predator acting as a blocker/herder most of the time. They also displayed more intelligent behavior in general, pausing more often rather than just blindly chasing the prey so that they could get more strategic positioning on the prey. In the trials that are shown in this video, the predators accomplished the following by the end of each trial:
This video is of the Both1M predator team, where the predators are given a combination of team and individual fitness functions and have unimodal neural networks. The color trails seen behind each predator track where the predators have been most recently. This team had a similar performance to that of the Individual1M team. The predators in this video display behaviors that appear to be a mixture of the Team1M and Individual1M teams. The Both1M predators act aggressively as a team as the Team1M predators did, but the Both1M predators have at least one predator who takes on more of a support role as a blocker/herder as the Individual1M predators did, but while still being aggressive. In the trials that are shown in this video, the predators accomplished the following by the end of each trial:
This video is of the Team2M predator team, where the predators are given only team-based fitness functions and have bimodal neural networks. The color trails seen behind each predator track the current module that the predator is using. The predators seem to use one module (the red module) when they are taking on aggressive behavior (selfish behavior) and they seem to use the other module (the light blue module) when they are taking on more of a support role as the blocker/herder (altruistic behaviors). This team performed better than any team with only one module, and was able to develop specializations that could switch between different predators at the appropriate times. In the trials that are shown in this video, the predators accomplished the following by the end of each trial:
This video is of the Individual2M predator team, where the predators are given only individual-based fitness functions and have bimodal neural networks. The color trails seen behind each predator track the current module that the predator is using. The predators used the two modules in the same way as Team2M did. This team performed better than any team with only one module, as well as slightly better than any two module team, and was also able to develop specializations that could switch between different predators at the appropriate times. This team also demonstrated behaviors that could be seen in the Individual1M team, such as more pausing and more patient behaviors, but did so in a more effect manner due to the addition of the second module. In the trials that are shown in this video, the predators accomplished the following by the end of each trial:
This video is of the Both2M predator team, where the predators are given a combination of team and individual fitness functions and have bimodal neural networks. The color trails seen behind each predator track the current module that the predator is using. The predators used the two modules in the same way as Team2M and Individual1M did. This team performed better than any team with only one module, as well as slightly better than the Team2M team, and was also able to develop specializations that could switch between different predators at the appropriate times. This team demonstrated a mixture of all behaviors, with a spread of aggression combined with support behaviors just like the Both1M team, except that these behaviors were more effective for the Both2M team due to the flexibility of the bimodal network. In the trials that are shown in this video, the predators accomplished the following by the end of each trial: