Agent Learning and Strategy Invasion
Main Menu Previous Topic Next Topic
Note: in order to run the simulation referred to in this slide, go here, to the Java Applet version. You will be directed to download the latest version of the Java plug-in.
In the graphic to the left, we present a simulation much like the ones in previous slides, but with several
1. there are a total of 20 agents, one of which is an All-D, and the rest are Tit-for-Tats;
2. at the end of each instance of Iterated PD, agents disconnect, and each connects to a random partner, not necessarily the one she was interacting with before;
3. after 6 instances of Iterated PD, each agent looks at every agent she has interacted with during these 6 games, and switches to the strategy of the one with the highest payoff, provided that it is higher than the agent's own total payoff.
Point 3 above is the crucial addition that reflects the concept of learning from the last slide. Every 6 Iterated PD games, a number we call generation length, each agent adopts the behavior of other agents she has seen as superior. Furthermore, all the payoff scores are set to zero at the end of a generation, and every agent is given a chance to start afresh. In such a simulation, the only thing that really matters is which strategies tend to thrive, and which dwindle to naught.
Pressing "Go" several times will have the agents play instances of Iterated PD with their partners, just like in the previous slides. At the end of 6 such instances, however, all the payoffs will suddenly go to 0, and the following step will reveal that the agents have decided to switch strategies: most likely, some will turn up to have a different color than that during the last generation. Several scenarios could happen every time due to the probabilistic nature of the game. For instance, the All-Defection strategy might start taking over parts of the society, invading it ideologically.
Previous Slide Next Slide