neural networks research group
Emergence of Competitive and Cooperative Behavior and Arms Race Through Coevolution (2010)
Author: Aditya Rawal, Padmini Rajagopalan
The following videos show an arms race sustained between coevolving predators and prey. A team of predator agents and a team of prey agents were coevolved, with the agents cooperating within the team and competing across teams. Such simultaneous competitive and cooperative coevolution can be used to construct complex behaviors in both predators and prey, each team alternately succeeding in an arms race. The teams had a Multi-Component ESP architecture, where each agent consisted of multiple neural networks to sense the opposing team members. The outputs of these neural networks were then given to a combiner network that decided the next move of the agent. The weights of all these networks were evolved. The fitness of the whole predator or prey team was then distributed equally among all the team members and thus, among all component neural networks.
A toroidal grid world was used to evaluate the predator and prey teams. There were three predators in all the simulations, and the number of prey was either one or two. The prey moved at the same speed as the predators in all the simulations. Each agent could move in four directions (east, west, north, south), and all the agents in the simulation made one move simultaneously at every time step.
In the videos below, the colored cubes are predators and the black spheres are prey.
Single Prey Agent
Phase 1: Predators Catch Prey
Initially in generations 50-75, the prey evolves only a greedy fleeing strategy, where it moves away from the closest approaching predator. Simple predator behavior is enough to catch the prey in this case. Two predators block the prey and the third approaches it from the third direction.
Phase 2: Prey Evades Predator
At generations 75-100, the prey evolves to selectively use the option of fleeing from the closest predator, and sometimes goes around in a small circle with the closest predator following on its tail. At this stage, the other predators too move between fixed positions without making any new move to catch the prey because they are acting as blockers.
Phase 3: Two Predators Cooperate
For generations 100-150, the predators learn to avoid the above deadlock. Two of them now approach the prey from opposing directions (acting as attackers) and the third one assumes the role of blocking.
Phase 4: Prey Baits and Escapes
In generations 150-180, the prey demonstrates intelligent baiting behavior by waiting for the two predators to converge towards it before moving away in a direction opposite to that of the predators. Since the third predator, the blocker, remains mostly stationary, the prey can easily dodge it.
Phase 5: Predators Switch Roles
To counter this move, the predators learn to dynamically switch roles in generations 180-200. The blocker also starts to follow the prey when it tries to escape.
Two Prey Agents
All of the behaviors that evolved in the single prey scenario also evolve here. In addition, a few more behaviors also emerge.
Predators Herd Prey
After some time, predators are able to succeed in herding the prey and capturing them simultaneously.
To counter herding, the prey evolve to scatter in different directions just before the predators converge on them. One reason for this last-minute scattering could be that once the predators have almost converged, they are all roughly in the same location, making it easier for the prey to evade them.
Kay E. Holekamp
holekamp [at] msu edu
Coevolution of Competitive and Cooperative Agent Behavior
Learning Strategic Behavior in Sequential Decision Tasks
2009 - 2014
Constructing Competitive and Cooperative Agent Behavior Using Coevolution
Aditya Rawal, Padmini Rajagopalan and Risto Miikkulainen
IEEE Conference on Computational Intelligence and Games (CIG 2010)
, Copenhagen, Denmark, A...
Coevolution of Role-Based Cooperation in Multi-Agent Systems
Chern Han Yong and Risto Miikkulainen
IEEE Transactions on Autonomous Mental Development
, 1:170--186, 2010.
IJCNN-2013 Tutorial on Evolution of Neural Networks
To Appear In 2013. Tutorial slides..
Areas of Interest