4.1 The Set-up

4.1.1 The environment

In the crowd-avoiding scheme each agent gets the most utility for going when less than 7 of the other agents go (0.7), they get a fixed utility (0.5) if they do not go and the lowest utility for going when it is crowded (0.4). In this way there is no fixed reward for any particular action because the utility gained from going depends on whether too many other agents also go. In this way there is also no fixed goal for the agent's learning, but it is relative to the other agent's behaviour (which will, of course, change over time). Under this scheme it is in each agent's interest to discoordinate their action with the others (or, at least, a majority of the others).

The friendly scheme is similar to the crowd-avoiding scheme, there is a basic utility of 0.5 for going if it is not crowded, and 0.2 if it is but if they go to the bar each agent gets a bonus (0.2) for each `friend' that also goes. If they stay at home they are guaranteed a utility of 0.65, so it is worth going if you go when it is not crowded with at least one other friend or if it is crowded with 3 or more friends. Who is a friend of whom is decided randomly at the beginning and remains fixed thereafter. Friendship is commutative, that is if A is a friend of B then B is a friend of A. An example of such a network is illustrated in figure 4. The number of friendships and agents is constant across runs but the detailed structure differs. In this scheme it is in the interest of agents to go when their other friends and only their friends are going. Under this scheme it is in each agent's interest to coordinate its actions with its designated friends but to discoordinate its action with the other agents.

Figure 4. An imposed friendship network

Under both schemes it is impossible for all agents to gain the maximum utility, there is always some conflict to provide a potential for continual dynamics.

Generated with CERN WebMaker