AGENT WORLD IN ECONOMIC LEARNING

摘要

Holland's original classifier systems fall victim to the "tragedy of the commons" due to their credit assignment mechanisms. Since it is possible for more than one agent to be receiving credit from the world at any given time, agents will pop up to receive credit even if they perform no useful service to the world. We can solve this problem by assigning ownership to the world to one agent at a time. Access to the world is now a commodity which can be bought and sold by competing agents, making credit and blame assignment trivial. By looking on the meta-learning level, economic agents with very simple bidding rules can work in cooperation to solve abstract complex problems.