On the influence of learning time on evolutionary online learning of cooperative behavior
Jörg Denzinger, Michael Kordt · 2001
We present an online learning approach for learning cooperative behavior in multi-agent systems based on invoking an offline learning method as a special action learn. We apply this approach to evolutionary offline learning using situation-action-pairs and the nearest-neighbor rule as agent architecture. For the application Pursuit Games we show that the online approach using evolutionary offline learning allows for good success rates for rather different game variants. Particularly, we perform experiments highlighting the influence of the time needed for learning and of the parameters of the evolutionary offline method. Our results show that even a duration of learn which is several times longer than the usual duration of an agent's actions still achieves good success rates. The same applies to rather small values for the key parameters of the offline method. Together, this suggests that this evolutionary online learning approach is a very good alternative to the well-known online approaches based on reinforcement learning.