So - given that people interact with each other in a social manner, to what extent are these interactions governed by the rules of their social group? I don't think it's to a very large extent. Look at honeybees - they interact with each other, but there isn't much of what I would call social interaction - they don't negotiate for greater status within the hive, they just do what they're wired to do. People by contrast still interact a great deal. I think that social groups provide a set of short-circuited interactions. They can also provide a forum for other interactions to take place. For example, within a social group, it may be pre-negotiated that people won't use violence to demonstrate social superiority, but that doesn't mean people won't jockey for position within the heirarchy, they'll just use a different method.
A hierarchy is useful in preventing conflict, and rapidly choosing someone to coordinate action when necessary. For example, suppose that there is a group of agents in an environment. It is accepted that these agents are amoral, and so they will not have any qualms about attacking each other to steal the other guys food/money/points/glass beads/whatever. However, a fight is pretty inefficient. Both parties may be hurt. So we can start with a Bullies relationship. If one agent is stronger than another agent, they will demonstrate this by bullying them. This is can be done by lower intensity conflict in a situation when no resources are at stake. This allows both parties to learn (or reinforces what they already know) who is stronger. Then when the weaker member of the relationship has some resources, and the stronger member wants them, the weaker member knows enough to give up without a fight. Note that this isn't always the way it would go - sometimes the weaker party will need the resource badly enough that it's worth fighting despite the risk. A more complicated social hierarchy would have the weaker party giving over some portion of his resources, depending on what he thought the odds would be if they fought. If the stronger party disagrees, he can initiate a fight anyway. If he agrees, he gets some free resources without any risk. A hierarchy is essentially a chain of Bully relationships in a situation where Bullying is transitive. It seems cruel, but consider this example: A tribe of cavemen, where everyone hunts for food within the tribes hunting grounds. Obviously the cavemen need to eat, and the more they can eat, the better off they are. In a bad year, some cavemen are going to die. You can minimize this number by allocating more food to the healthier cavemen. A hierarchy accomplishes this because the healthier cavemen are the ones at the top of the hierarchy.
Hierarchies can also be useful in less sever circumstances. You can use them to avoid arguments by a similar means. For example, suppose a chess club, with a hierarchy based on who can be who in chess. Essentially, all a hierarchy does is propogate useful information (relative strength) throughout a social group. In a chess club, it's mostly used to avoid unbalanced matches, which aren't much fun for either player. (I can't think of any other uses in this particular example right now, but there probably are). Note that Hierarchies are efficient about this. It's somewhat similar to the operation of a sorting algorithm (now theres an interesting idea)
Given the existance of a set of rules for some overarching group, the next issue is how and what kinds of interactions are supported by the rules of the group. In one sense, starting at the group level is impractical, but necessary. Furthermore, I suspect that people have some build in interaction rules, like vengeance. (When you think about it, revenge isn't practical, but it can be useful in a social context, in terms of dissuading others from attacking you.)
Just to summarize what I'm trying to do. If you're familiar with the Survivor TV show, you will be aware that it features an arena of competition which, as far as I know, hasn't been looked at too closely in AI - the problem of Social competition. I believe that there are a number of strategies used by humans to deal with that problem, and I think they could be adapted for use by a computer. I'm reading Reasoning about Rational Agents right now, and while I think that book is very useful, there are some problems that it doesn't confront. For instance - when you are in an environment with other agents which are roughly as complex as yourself, it is computationally impossible for you to exactly predict what their actions will be. Since you can't predict, you will have to guess. But people (and therefore, presumably, computer agents) don't behave randomly. I believe that people use social rules to predict what other people are going to do (frex: I know than Al does not like Bob, therefore I can predict that Al will not take an action which helps Bob). So I'd like to set up a formal description of social interactions which computers can use to make predictions of that type.
In order to do this, I need to set up an environment in which computer programs can interact competitively with humans. This is particularly difficult, as there are a number of self-contradictory properties I'd like the environment to possess. For ease of analysis, I'd like the environment to be expressed in terms of normal form games. However, it is also clearly necessary for agents (programs or human players) to be able to communicate with each other, which makes it pretty much impossible to describe it completely as a normal form game. It's necessary for agents to be able to interact cooperatively or competitively, and I'd also like it to be possible for an agent to try to avoid interacting with another agent. It seems to me that the best way to do this is to impose some sort of location rules on the game, but I'm not sure. I would prefer to avoid an absolute utility function for the game, which is somewhat contrary to the idea of expressing everything in terms of a normal form game. This makes the problem of assigning goals to an agent fairly difficult too. I'd like to include the idea of objects or tokens which can be exchanged as well. Furthermore, there needs to be a concept of duration - time for the agents to interact in. Strict normal form games aren't interesting unless played iteratively. I've played a lot of board games anyway - I'll have to look at them for ideas.
The views and opinions expressed in this page are strictly those of the page author.
The contents of this page have not been reviewed or approved by the University of Minnesota.