Initial Experiment

In exploratory experiments [43], we used a Genetic Algorithm to learn the weights of a perceptron network that played Tron. We found that simple agents played interestingly, but also that coevolution may not always lead to robust strategies. Collusion [107] was likely to appear in the form of ``live and let live'' strategies such as the one shown in figure 3.3, where agents avoid confrontation and ``agree'' to tie, splitting the point.

  

Figure 3.3: Live and let live: Two artificial Tron players make tight spirals in order to stay as far from the opponent as possible. This form of collusion is a frequent suboptimal equilibrium that complicates artificial learning through self-play.