Dynamic Pricing Agents and Multiagent Learning.pdfVIP

Dynamic Pricing Agents and Multiagent Learning Junling Hu University of Rochester January 15, 2003 Abstract. We implemented three different types of pricing agents in a simulated economy. Each type of agent is based on a different learning method. The first method is simple rein- forcement learning. The second method is the traditional Q-learning method. The third method is Nash Q-learning method. In each simulation, there are two agents, and a fixed amount of customers. The agent that charges a lower price will attract all the customers. When both firms charge the same price, they receive equal share of the market. Our simulation shows that the two Q-learning methods that take future rewards into account perform better than the simple method that is myopic. Among the two Q-learning methods, Nash Q-learning performs con- sistently better than the Q-learning method. This suggests the importance of game theoretical modeling in online learning. Keywords: Price agents, multiagent systems 1. Introduction The presence of software agents on the Internet has dramatically changed the way stores compete with each other. With search engine and compari- son shopping bots gathering prices from different stores instantly, any price difference among stores will be instantly detected by the customer. Even though there exists some variance among store’s prices, such variance has been shrinking. Is there a learning method that allows the seller to consistently perform better than other methods? Another question we are concerned is: Would the economy with pricing agents lead to more or less volatile markets? What kind of price equilibrium would the market converge to under a pricing strategy? In this paper, we investigate how a software agent can help online stores monitor other stores’ prices continuously, and then set corresponding prices. Such agents are called pricebots (Kepha