Let’s Make a (Virtual) Deal

Researchers at USC’s Institute for Creative Technologies have created human-like software agents to help people become better negotiators

Years ago, when Jonathan Gratch went to buy his first car, he became so distressed about the idea of negotiating the price that he bought a Saturn.

The carmaker had a no-haggle price policy that was ideal for skittish consumers. Like many car buyers, Gratch dreaded having to engage in the back-and-forth of trying to secure the best deal.

Gratch, a computer science research professor at USC Viterbi with a dual appointment in psychology, still abhors negotiating, and considers himself bad at it. But as director of Virtual Humans Research at the USC Institute for Creative Technologies, he’s in a position to do something about it.

Gratch and a team of four Ph.D. students in computer science, along with a programmer and a research scientist, have been working on a project called the Conflict Resolution Agent. They hope to have this virtual negotiator ready sometime this year.

The “agent” is a virtual human that has been fed reams of information gleaned from interactions between 300 real people who, in pairs, haggled over items in a scenario Gratch created based on the TV show “Storage Wars.”

In Gratch’s version, called “Antique Wars,” two people are given instructions for which items they want — lamps, old records, an Art Deco print — without knowing what their partner wants. The participants, recruited from Craigslist and given a monetary incentive to do their best to get what they want, negotiate for 15 minutes.

Gratch and his team videotaped the negotiations, cataloguing what the participants said and noting their facial expressions and body language. Using that information, the team is exploring two parallel paths that would allow people to negotiate with a virtual human.

One of those is the Conflict Resolution Agent, an avatar designed so a person can interact with it “face to face” via spoken language. It most likely will run on a desktop computer initially.

The other is an online platform called IAGO, for Interactive Arbitration Guide Online. IAGO, developed by Johnathan Mell, 25, a fourth-year computer science Ph.D. student, is intended to be an automated agent that a person can interact with via menu-driven interaction over the web.

Unlike the Conflict Resolution Agent, IAGO doesn’t use spoken language and won’t sense a person’s nonverbal behavior, but it can reach a larger audience. Gratch’s team is opening up IAGO to other researchers as part of a competition to create intelligent negotiators.

Teaching a virtual agent how to negotiate was a time-consuming process. The team programmed the virtual agent to say about 5,000 distinct utterances that the real people said when they faced off in the “Antique Wars” scenario.

At this point in the development of the virtual agent, a human is needed to operate what the agent says and how it reacts. The operator punches buttons that control the responses the virtual agent makes based on what the real person says in the two-person negotiating process. One button may have 20 variations of the same statement.

“It’s kind of like learning to play a musical instrument,” said research analyst Sharon Mozgai, a member of Gratch’s team who controls the Conflict Resolution Agent’s “brains.” “The human who controls the virtual agent figures out how to play this keyboard that makes the virtual character talk in a certain way.”

The technology is developing so fast, however, that Gratch anticipates having a beta version of a fully automated virtual negotiator this year — an agent that requires no human to interact with a real person. A version of the agent exists, but “it’s kind of dumb,” Gratch said with a chuckle.

“Most people typically are bad at negotiating,” Gratch said. “They find it unpleasant and feel it’s confrontational. And companies spend billions of dollars a year hiring consultants to train their employees to become better negotiators.”

Using virtual negotiators to train humans will expand the fine art of negotiation beyond the realm of the elite, that is, those who can afford expensive training, Gratch said. This is important for several reasons.

For one, there is the gender pay gap. Studies show that women are more reluctant to negotiate for higher salaries than men. Providing women an online platform to become better negotiators by practicing with a virtual human could help trim that gap, Gratch said.

It also could have implications in the area of civil litigation, where 90 percent of cases end up in negotiated settlements. Having more lawyers better trained in negotiation could yield more equitable settlements.

Although artificial intelligence has been around since the early 1950s and computers now can do many things amazingly well, like playing a movie or turning on the lights after prompted by a voice command, researchers only are beginning to scratch the surface when it comes to teaching computers soft skills like negotiating and becoming more emotionally intelligent, Gratch said.

“From the outside,” he said, “the technology seems super amazing, but from the inside, you can see all the limitations.”

ICT is unique, Gratch said, because the center brings together experts in such diverse fields as natural language, graphics and computer animation, modeling cognition, modeling emotion, cinema and even medicine.

Gratch has co-written a proposal with David DeVault, a colleague in the Department of Computer Sciences, for a grant from the National Science Foundation to take the virtual negotiator to the next level. If approved, Gratch plans to test the technology in the negotiation class of Peter Kim, a USC Marshall professor of management and organization.

Kim said a potential benefit of the technology is that it could help overcome a common limitation in how negotiation cases are typically implemented and experienced. This usually takes the form of having students complete a case just once, with the opportunity to make just one sequence of decisions as they proceed through that experience.

“As a result,” Kim said, “it’s natural for students to wonder what would have happened if they tried something different, how much of their own success was based on their own behavior or the idiosyncrasies of their counterpart, and how alternative strategies would have affected the ultimate outcome of the negotiation.

“The ICT system,” Kim said, “could help overcome this issue to the extent that it can be adapted in a way that would allow students to try the same negotiation multiple times with the same virtual counterpart, and thus gauge the implications of approaching the case differently.”

Right now, the focus of the research being conducted by Gratch’s team has been on getting the language to work so the virtual agent can build up a vocabulary for the negotiating process. The following is an exchange between a real person and the Conflict Resolution Agent as they haggled during an “Antique Wars” scenario:

(Agent) That doesn’t seem fair, though.

(Person) It doesn’t?

(A) Not really.

(P) I guess we can’t come to an agreement.

(A) What do you think about me getting two chairs and the clock and you getting one chair and both plates? You’ll probably make most of your profits off of these plates.

(P) No, I won’t. Because I know these chairs have more value.

(A) Yeah.

(P) So I will not let you have two chairs.

(A) Of course the chairs are worth something.

(P) Yes, they are, so I’m going to take all three.

(A) That doesn’t seem fair, though.

Mell’s focus is on creating virtual agents with distinct abilities, such as one who is good at lying, a classic negotiating tactic, and another who readily volunteers information.

“I’ve always been fascinated by the idea of creating really detailed characters that can act in a human-like way,” said Mell. “Most people have this impression that computers are robotic, and that’s a pejorative term: ‘Oh, they can’t understand a lot of things.’ So we’re trying to make that not be quite as true.”

“A lot of the research we’ve done up to this point is one on one and structured,” he added. “I’m focused on the long term on such issues as, how can we create a virtual agent who can interact with a human in a more free-form fashion?”

Bottom line, Gratch sees the Conflict Resolution Agent as a valuable way for more people, from students to lawyers to business people and beyond, to put into practice what they may have learned from textbooks, classroom lectures and seminars.

“Knowing is not doing,” Gratch said. “You can know a lot about negotiations but still be really bad at it.”

The professor may be first in line to practice with a fully automated virtual negotiator. “When I bought my most recent car a few years ago,” Gratch explained, “the salesmen did the good cop/bad cop strategy on me.”

Gratch knew exactly what they were doing intellectually.

“It was a technique to establish a sense of positive emotion and alliance that could be subsequently exploited,” he said.

It worked.

“I paid more for the car than I had hoped,” Gratch said with a smile.