Using GDL to Represent Domain Knowledge for Automated Negotiations Dave de Jonge Western Sydney University Sydney, New South Wales, Australia d.dejonge@westernsydney.edu.au Dongmo Zhang Western Sydney University Sydney, New South Wales, Australia d.zhang@westernsydney.edu.au ABSTRACT Current negotiation algorithms often assume that utility has an explicit representation as a function over the set of pos- sible deals and that for any deal its utility value can be calculated easily. We argue however, that a more realistic model of negotiations would be one in which the negotiator has certain knowledge about the domain and must reason with this knowledge in order to determine the value of a deal, which is time-consuming. We propose to use Game Description Language to model such negotiation scenarios, because this may enable us to apply existing techniques from General Game Playing to implement domain-independent, reasoning, negotiation algorithms. Keywords Automated Negotiation; General Game Playing; Game De- scription Language 1. INTRODUCTION Most work on Automated Negotiations focuses purely on the strategy to determine which deals to propose given the utility values of the possible deals. Little attention has been given to negotiation settings in which determining the utility value of a deal is itself a hard problem that takes a substan- tial amount of time. One often assumes the utility value of any deal is known instantaneously, or can be determined by solving a simple linear equation [1]. In such studies the process of evaluating the proposal is almost completely ab- stracted away and one either assumes that the negotiation algorithms do not require any domain knowledge or reason- ing at all, or that all such knowledge is hardcoded in the algorithm. The preferences of the agent’s opponents on the other hand, are often assumed to be completely unknown. In this paper however, we argue that in real negotiations it is very important to have domain knowledge, and a good negotiator must be able to reason about this knowledge. One cannot, for example, expect to make profitable deals in the antique business if one does not have extensive knowledge of ACM ISBN 978-1-4503-2138-9. DOI: 10.1145/1235 antique, no matter how good one is at bargaining. Moreover, a good negotiator should also be able to reason about the desires of its opponents. A good car salesman for example would try to find out what type of car would best suit his client’s needs, in order to increase the chances of coming to a successful deal. We therefore propose a new kind of negotiation setting in which the agents do not have an explicit representation of their utility functions but instead are presented with domain knowledge in the form of a logic program. Agents will need to apply logical reasoning in order to determine the value of any proposal. Another point that is rarely taken into account, is that an agent’s utility may not always solely depend on the agree- ments it makes, but may also depend on decisions taken outside the negotiation thread. For example, suppose that you negotiate with a car salesman to buy a car. If you are single and you live in the city then it may be a very good deal to buy a small car which is easy to park and uses little fuel. However, if one year later you get married and decide to start a family, that deal suddenly is not so good anymore because you now require a larger family car. Interestingly, we see that although the deal itself has not changed at all, its utility value certainly has changed as a consequence of some decision taken long after the negotiations had finished. Moreover, an agent’s utility may not only depend on its own actions, but also on actions of other agents, as is typical for business deals. Imagine for example renting a property to open a restaurant in a street with no other restaurants. This might be a good deal until suddenly five other restau- rants also open in that same street, giving you so much competition that you can no longer afford the rent. We note that these properties we are addressing here– applying logical reasoning about the domain, and choosing a proper strategy with respect to your opponents’ strategies– are also the main issues in the field of General Game Playing (GGP). General Game Playing deals with the implementa- tion of agents that can play any kind of game. In contrast to specialized Chess- or Go- computers, which can only play one specific game and are largely based on knowledge pro- vided by human experts, a GGP program cannot apply any game-specific heuristics because it only knows the rules of the games it is playing at run-time. Therefore, in this paper we propose to use Game Descrip- tion Language (GDL), which is commonly regarded as the standard language for GGP research [18], to define negoti- ation domains and we propose to use common techniques from GGP to implement negotiating agents. We investigate