Skip to main content
European Commission logo print header

Information Exchange Policies for Human-Computer Negotiation

Final Report Summary - IEPHCN (Information Exchange Policies for Human-Computer Negotiation)

This report summarizes four years of research efforts on the project Information Exchange Policies for Human-Computer Negotiation (IEPHCN num 268362).
Our work focused on the development of computational models for human-computer negotiation in strategic settings, and the sue of information exchange strategies for constructing systems that are able to negotiate successfully with people.

Our research efforts in these years have focussed broadly on the following thrusts:
1. Modeling people’s information revelation behaviour in various strategic settings and in different cultures.
2. Decide whether and under what conditions to reveal information that other participants may not know and when to avoid revealing such information.
3. Learn and adapt to the reliability and trustworthiness of people over time.
4. Modeling people’s commitment behaviour in negotiation.
We have designed models for meeting the above objectives using a variety of techniques, including game theoretic reasoning, decision theoretic models, and machine learning. We have evaluated our proposed models in extensive experiments involving hundreds of human subjects and compared the results of our models to that of humans, as well as the state of the art strategies from the literature.

We details our main results from the study. First, we have designed several settings in which we study people’s information exchange behaviour. The settings were configured using Colored Trails, a multi purpose system for studying decision making in settings that comprise both people and computer agents. The use of these settings allow to model people’s decision-making and information exchange strategies in settings of varying complexity and in different cultures; and, the design of agents that are able to outperform people in these settings. The following lists the main results and impact from each of the thrusts.

Result 1: The design of novel settings for studying argumentation and negotiation. We designed three classes of strategic games. The first is called “revelation games” -- bilateral bargaining games in which agents can choose to truthfully reveal their private information before engaging in multiple rounds of negotiation.

The second setting is called “contract games” — a board game in which three bidders compete to win contracts by submitting bids in repeated auctions, and a single auctioneer determines the winner of each auction. Bidders earn bonus points in the game if their bids are accepted. The auctioneer's score is constant and does not depend on their bids. The winning bid is paid to an external “government” entity. Throughout the bidding process participants could send unlimited private messages to each other.

The third setting is called the Contract Game which is analogous to a market setting in which participants need to reach agreement and commit or renege from contracts over time in order to succeed. The game comprises three players, two service providers and one customer. The service providers compete to make repeated contract offers to the customer consisting of resource exchanges in the game. The customer can join and leave contracts at will.

Result 2: Constructing decision-making and information exchange models for agents that negotiate with people in strategic negotiation scenarios.
Results in this thrust were comprised of two interleaving efforts. The first is the design of game theoretic models of people’s information exchange strategies. To this end, we formally define the notion of commitment between service providers and customers in the game and provide Sub-game Nash equilibrium strategies for each of the players. Specifically, because service providers compete over the customer player, the contracts proposed by both service providers and customers are highly beneficial contracts to the customer, but require a commitment from the customer that would prevent it from signing a contract with the other service provider. These off-the-equilibrium path strategies are shown to be especially relevant to human play in the game which does not adhere to equilibrium strategies.

A third class of agents combined a decision-theoretic model with machine learning to generalize the agent to different people. The learning techniques were based on features that represent players' states in the negotiation, as well as social factors that reflect their generosity and reliability over time.

Another class of agent strategies used predefined rules to construct a model of its negotiation partners, and domain knowledge to guide the agent when searching for the best proposals to make. The advantage of this agent-design is that it does not rely on training data of human play and is able to adapt to individual people very quickly during the negotiation process. The disadvantage of this agent-design is that it requires hand-designed rules from a domain expert.

Result 3: Evaluation and empirical analysis.

All of the agent-designs were evaluated when negotiating with new people in different countries. We used hundreds of subjects for this purpose that were recruited from different countries and different cultures, spanning gender, ages and social economic status.
Specifically, n the contract game, we evaluated computer agents that use the equilibrium strategies, we conducted extensive empirical studies in three different countries, the U.S. Israel and China. We ran several configurations in which two human participants played a single agent participant in various role configurations in the game.

In the corruption game, we played 284 games in Israel, the U.S and China. We analyzed the logs that were generated by the games posthoc in each country. We measured two types of corrupt activities, one in which the auctioneer did not pick the bidder with the highest bid, and in which there was a distinct bribe that was transferred from a bidder to the auctioneer in return for getting chosen. Bribery was defined the case in which the auctioneer accepted or solicited resources from one of the bidders, who was subsequently chosen for the contract despite not submitting the highest bid.

In terms of average performance, the agent using predefined rules was able to negotiate as well as people across all countries. However, in the U.S. alone, the agent outperformed people, while in Lebanon alone, the agent was outperformed by people. The learned based models were able to outperform people in all of three countries. In addition, they adapted to the changing behavior of their negotiation partners over time in each culture despite the limited amount of training data.

We found that in the contract game, the computer agent using Nash equilibrium strategies for the customer role was able to outperform people playing the same role in all three countries. In particular, the customer agent made significantly more commitment type proposals than people, and requested significantly more chips from service providers than did people. Also, the customer agent was able to reach the goal significantly more often than people.
Interestingly, we saw differences in the performance of the provider agents playing Nash equilibrium strategies between the different countries. Specifically, in China, people were able to outperform the provider agent, while in Israel the performance of the service provider agent was similar to that of people.

We found that in the corruption game, corruption occurred in about 33\% of games played in the U.S. 29\% of games played in Israel and 46\% in the games played in China. These results partially follow the perceived corruption index, in that China was the country that exhibited the largest amount of corruption. However, we measured lower level of corruption in the in Israel as compared to the U.S. despite the higher perceived corruption level in Israel. In Israel the presence of bribery benefited the auctioneer when compared to the case in which there was no bribery (the auctioneer chose the highest bidder). However, while the bribing players benefited from this activity in Israel and in China, they did not benefit from bribery in the U.S. Further analysis revealed that bribes in the U.S. often were excessively large and were more likely to prevent the bidders from reaching the goal. In all countries, we found that there was a significant increase in the ratio of corruption in both countries in the last round of play.

Together, these results provide significant contributions towards understanding human-computer decision making in strategic setting. His paper makes three major contributions. First, we provide a new empirical framework to investigate negotiation and information exchange mechanisms in a lab setting without priming subjects. Second, we show that it is possible to model people’s play in these strategic settings despite the inherent uncertainty that governs their behaviour as well as the fact that people do not adhere to retinal, game theoretic strategies. Third, we provide agent designs that can outperform people in these settings, while also facilitating their decision-making.