Robot Ecosystems

Towards A Common Evolution:
Cooperation Is Harder Than Competition

ASADA--Recently we've been studying genetic programming,[*13] particularly to what extent two agents are capable of cooperative behavior. The problem is that they aren't very cooperative with each other. With one driving force of will the concept of efficiency is usable, but with multiple will mechanisms the whole idea of efficient, coordinated action just falls apart. Figuring out how to let them work together to fool an opponent is also extremely difficult to manage. Two robots without a common enemy won't cooperate. There's no apparent necessity for them. Of course, it's more efficient for them to cooperate, but in defining priorities for their tasks, the tasks that they must divide between themselves, there are a lot of managerial adjustments that we must make or collaboration is really difficult to achieve.
Take passing and shooting for example. The robot assigned to passing's job is far more complex than the robot assigned to shooting. Not only do they have to pass accurately to the shooting robot, but if the shooting robot doesn't make their goal the pass is not appraised as having fulfilled its function. The result is that the passing robot would rather shoot the ball himself. It's easier that way. (laughs) It's easier when there are competitors on the field, but still, meaningful cooperation is difficult to achieve.

SAKURA--That's fascinating.

ASADA--I mentioned it at some research group recently and someone tried the experiment with their kid's junior league! There was one goalkeeper, and two other children who were supposed to be passing and shooting, but the goalkeeper was just in a far stronger position. It's easy to impede, tough to cooperate. In their game the keeper was always winning, while the attackers were losing spirit, and the kids were losing interest in the game. It was only after they put more restrictions on the goalkeeper's actions that they all started to coevolve.

SAKURA--There is a British psychologist who conducted a classic experiment with a dominant pig and a subordinate pig, where the object was to push a button in one location to make food come out of a dispenser in another.
Against expectation, the pigs did not emphasize difference, but developed a cooperative relationship. The dominant pig would push the button, and the subordinate pig would sit, immobile in front of the food dispenser. Once the dominant pig had pressed the button enough, it would come over, and chase the subordinate pig away, finishing what was left. The conclusion that this psychologist made from their behavior was that they did this because the subordinate pig would have been helpless on his own. It would have been in the way of the button, and therefore the dominant pig's capacity for getting food. With no other viable alternatives for it, it would press enough so that even after the subordinate pig had eaten some there would be enough left over. It was a reasonable compromise for both parties.
Of course, the results of the exercise would have been different had there been a different amount of food available, or if the button and food dispenser had been differently located, so this cooperation was not limitless, but it was demonstrated that these pigs were able to demonstrate a cooperative arrangement appropriate to their physical capacities.

ASADA--That is precisely the kind of problem that is needed to organize complex relationships and synchronize behavioral patterns for coevolution.

SAKURA--I suppose so. With human beings, societal rules function as such external guiding principles. Of course, humans may be one animal that is inherently good at cooperation, but I suppose that this is the result of a long road of conflict with various external groups. For example, groups that found plentiful water supplies were bound together to protect it. External pressures. The existence of a third party is what forwards the body politic.

ASADA--I think that it's interesting that competition is innate, but collaboration is learned. Games don't often stress such situations. We've finally gotten the lower level behaviors in line, the shooters to learn shooting, but we haven't yet been able to enable the higher levels of cooperation. What we've done is to simulate human society, where in this case you would have a coach working with the team to encourage coevolution,[*14] "bottom-up" methodologies. Even though they may not learn a given skill the first time it is taught to them, after repeating the skill along with the coach several times, they begin to grasp what was being explained to them. Also, the coaches learn how to express themselves better over time. It is an experiment in collaborative development. Rather than all of the robots beginning from the same starting point, some, which are designated as coaches, are able to interact with the player robots and they both learn how to function within their roles together.

Backleft rightNext