Learning in a market system


A real economic system provides monetary incentives to its participants and dynamically allocates the available system resources in part because the human beings and corporations that take part in it are smart. They can recognize when they are charging too little or too much, or when they are not getting the quality they expected. They also do not spend all their time thinking about the economy, but only do as much strategic thinking as is needed.

If we want our agent economy to be as robust as the real economy we will need to have at least marginally intelligent agents. These agents will need to know what to bid-- both when price has reached an equilibrium (which is easy), and when the price is fluctuating. If an agent is the only seller of a service then it should be able to take advantage of its monopoly, while if the buyers find that a seller's prices are too high for the service it sells, they should be able to avoid buying from him. Agents will also need to have enough computational power left over to actually deliver their service.

Using learning agents frees us from having to implemented some sort of centralized ``police'' agents. While the UMDL ontology provides a way for agents to characterize the services they sell, there is no guarantee that all the goods sold at an auction for service x are indeed instances of service x. This guarantee could be provided by police agents that periodically check all the goods sold at all the auctions. Unfortunately, this would a) be computationally taxing solution on the system, b) give rise to thorny political problems given the fact that the police agents would have special rights over other agents and c) some agents might have specific subjective preferences over a good which can not be expressed in the ontology (and, therefore, not recognized by other agents). For example, while all agents might agree that QPA1 does sell service x, one agent might think that QPA1's service is faster, better, or more thorough than the same service x as provided by QPA2. This being the agent's subjective opinion, it is unlikely to be in agreement with all the other agents, but the agent might still be willing to pay more for service x from QPA1 than from QPA2. Giving agents the ability to learn is also a first step towards the implementation of ``recommender'' agents that gather together agents of similar tastes.

In essence, learning provides a way for agents to build trust among each other. For a market-system to work well, an agent needs to be able to trust that the other agent's view of good x is the same as his view of x. Similarly, an agent that uses recommender agents needs to trust the recommendations they give. This trust can be acquired by repeated iterations with the agents in question. Once the trust is acquired the learning is no longer needed, that is, until the trust is broken. This is why we argue that agents need the capability of learning, even if this capability is not always exercised.

Lastly, we propose that learning agents are not only useful, they are inevitable. In a society of selfish agents, we can expect that the designers will use every technology available to enhance the profits of their agents. Learning is one such technique. By implementing learning agents ourselves we can determine how much of an advantage they will have and how they will affect the system.

Jose M. Vidal
Tue Sep 30 14:35:40 EDT 1997