Opponent Modeling for Virtual Human Negotiators

Abstract

Abstract. Negotiation is a challenging domain for virtual human research. One aspect of this problem, known as opponent modeling, is discovering what the other party wants from the negotiation. Research in automated negotiation has yielded a number opponent modeling techniques but we show that these methods do not easily transfer to human-agent settings. We propose a more effective heuristic for inferring preferences both from a negotiator's pattern of offers and verbal statements about their preferences. This method has the added advantage that it can detect negotiators that lie about their preferences. We discuss several ways the method can enhance the capabilities of a virtual human negotiator.

The Impact of Implicit Information Exchange in Human-agent Negotiations

“…Agents were more accurate in inferring participant goals when they combined information from both the participant's statements and their sequence of offers. This reinforces but also extends the findings of Nazari and colleagues [11]. Nazari showed this accuracy can be achieved by analyzing negotiation traces after the fact.…”

“…First, we extend its opponent modeling methods to attend to implicit as well as explicit information. Specifically, we incorporate an opponent modeling method by Nazari that has been shown to more accurately model human goals [11]. Second, we extend its offer-generation method to implicitly convey the agent's goals.…”

“…Prior research suggests that agent perspective-taking is enhanced by attending to the information implicit in the human's pattern of offers. Nazari and colleagues examined the accuracy of different opponent modeling techniques at predicting human goals in human-human negotiations and found the most accurate models utilized both implicit and explicit information [11]. However, these models adopted a third-party perspective (they analyzed traces from completed negotiations).…”

“…To manipulate perspective-taking sophistication we replaced the default modeling behavior of Pinocchio with a set of "frequency models" proposed by Nazari [36]. Frequency models are a heuristic method for learning what an opponent wants by counting the frequency of certain events, such as how often someone makes a concession on an issue or how often they explicitly assert a preference for that issue (see an overview in [19]).…”

“…A core principle of negotiation is perspective-taking, or what is sometimes called opponent modeling [10,11]. By better understanding the goals of one's opponent, it is often possible to discover creative solutions that benefit both parties.…”