Generic modeling is our first attempt at opponent modeling. It assumes that our opponents use a value scale similar to our own. The observed actions of each opponent are used to adjust their weight array. No statistics are gathered, so the re-weighting system treats all actions equally, dependent only on the context and regardless of which player it is.
We pitted together four different versions of Loki for 100,000 trials. The focus of the experiment was four copies of GOM (a player using all betting strategy features, generic opponent modeling, and a tightness setting of loose). The competition was three different non-modeling players (two copies of each, also using all betting strategy features):
Figure 8.3 shows the results of the experiment. GOM quickly demonstrated clear dominance while the three non-modeling players were ranked based on the tightness of their play. As expected, GOM is able to exploit the basic players because its model of how they play is fairly accurate, and is used to make better decisions. GOM might not perform as well against players with very different styles of play, because its model would be less accurate, but it would be better than using no modeling at all.
