The Emerging Paradigm of Intelligent Negotiations 

Reflections on “BiLAT: A Game-Based Environment for Practicing Negotiation in a Cultural Context”

“Let us move from the era of confrontation to the era of negotiation”, as said  by Richard Milhous Nixon, the 37th President of the United States, gives us a brief idea about the importance of the art of negotiation in the era we are living in. Whether it’s a matter of professional life or personal life, a negotiation can always bring solutions without having to enter a contentious state. Apart from individual perspectives, it plays a very important role in national decisions as well. One of the very prominent areas of usage is our defense system. Military units of different countries are frequently found to be using their negotiation skills at various places to avoid conflicts wherever they can to minimize the harm caused by a battle. So, it often becomes crucial to train the soldiers or other leaders to have equipped with some of the best negotiation skills and hence the defense research is putting a huge effort and investment on the same. Intelligent Tutoring Systems(ITS) are smart automated tutors used to train people in various domains with the help of embedded Artificial Intelligence. One such system Enhanced Learning Environments with Creative Technologies (ELECT) Bi-Lateral Negotiation (BiLAT) is a prototype developed by The USC Institute for Creative Technologies (ICT), in collaboration with organizations within the U.S. Army to address the problem of negotiation skill training through the use of automated learning. In the aforementioned paper in the subtitle, the authors give a detailed overview of the basics of negotiation and the abstract overview of the BiLAT system and what its components are. The domain of negotiation from the point of view of Artificial Intelligence is very ill-defined which causes the designers of such ITS to handle many complexities associated with it. The description of the internal dynamics of negotiations in the paper are very comprehensive and gives the reader a very good knowledge about the domain. However, the details of the BiLAT system is completely verbose and doesn’t include any implementation level details which some readers like me may find little fuzzy if they crave to know more about “how” the different stages or modules were actually put into existence and it’s relevant technologies. The paper is concluded with a comprehensive description of the experiments and the results and we believe that it bolsters the work from a strong mathematical standpoint.

Breaking down the domain of Negotiations:

As mentioned earlier, the ill-defined domain of negotiations pose some serious limitations while implementing intelligent systems for tutoring these skills. Lack of a definition of a proper feedback and the hazy notion of the correctness of moves makes it even more difficult to evaluate the students. Most of the negotiation training typically take place as a role-playing activity and BiLAT system also uses a similar strategy. While building the system, they first tried to find out the actual forces that drive a good negotiation and tired formalizing it as much as it could be done. Previous literature reviews and survey with the military negotiation experts were one of the main inputs to the knowledge base. Finally, they had converged to two facets of the results as stated by the authors, (1) principles that were important for exercising good judgment and decision-making, and (2) procedural steps that supported the principles and increased the chances of success. They categorized the processes used by the experts in three broad phases namely preparation prior to a negotiation, negotiation, and follow-up after a negotiation. The process was further broken down into 12 more stages, the details of which will be omitted here due to space constraints. Also, as the BiLAT game was primarily focussed on military training and loads of those took place in the Middle East area, a huge attention was given on the cultural adaptability aspects. Finally, the formalized stages were converted to some well-defined Learning Objectives(LO) which the tutoring module used for accurate feedback and proper guidance.

The Student’s Perspective and Behavior Models:

During the execution, the central character, i.e. the student(or the player) enters in negotiation meetings with different people in the virtual environment and chooses different actions to perform based on which it progresses through the storyline. The authors have very well described all the stages of the negotiation process and how it looks from the point of view of the player with the help of actual examples of conversations and screenshots from the BiLAT gameplay. The student prepares for the meeting by collecting information on the counterpart and filling up some sheets. During the meeting, the user can select different actions from the GUI of the negotiation environment based on his decisions on the next moves. Depending on his moves, the different matrices of the scene keeps changing, for example the trust meter. The player can also take part in the followup activities and also check if the counterpart followed through anything. There are two main submodules which control the overall behavior of the character while the negotiation is in progress. The first one is called the dialog model which controls the dialog dynamics in the scenario and analyzing down their syntax and semantics. The dialog model is implemented as a rule-based state machine, where the player states on different variables like trust and influence are constantly updated based on player actions. The authors described the details of the model from the point of view of Middle Eastern negotiation scenarios which is very relevant to BiLAT. The reader gets a clear idea how the concepts of the model map to the gameplay mechanics of BiLAT as they describe the steps with real gameplay example screenshots. Unlike the dialog modeling, the negation model used in BiLAT was a previously implemented negotiation architecture called PsychSim, which can be modeled as a social simulation tool. The designers of BiLAT mapped the problem of negotiation to the existing architecture of PsychSim. It describes the characters as entities and their corresponding feature space to create a meaningful behavioral model. In author’s words, PsychSim generates the behavior for these entities and provides explanations of the result in terms of each entity’s goals and beliefs. In a nutshell, the player interacts with the autonomous agents of PsychSim that has its own goals and various negotiation moves. Both dialog model and the negotiation model of the BiLAT system helped the designers to take advantage of the well-defined structures of the negotiation domain while compromising very little on the realistic aspects of it.

How BiLAT does it:

The system is implemented to simulate a virtual environment in which the player has to make certain decisions and move towards the discussion with the counterpart. The various player actions instigate the system to provide feedbacks to the player in terms of the changes in the various player parameters like trust. Learning Objectives(LO) define the modular goals for the tutoring system which define the things the ITS wants to deliver the player as taking home messages. The Coaching aspect of BiLAT tries to categorize the player actions as correct, incorrect or mixed depending on the appropriateness of the actions corresponding to a particular counterpart, scenario or learning objectives for that particular phase. For example, if a player directly rushes down to main business in the opening talk itself, then the coaching module marks it as incorrect as it violates the LO of patience. However, the decisions of when to deliver the coaching messages depend on the actual play sequence of the player and the state of the environment and is decided dynamically. The final important module of BiLAT is the reflective tutoring system which strives to implement the after action review(AAR) phase of the structured negotiation model described above. After the gameplay is over, the players are presented with a detailed feedback and analysis of their actions and the reflective tutoring module tries to point out the mistakes made by the player during the last play. This helps in effectively increasing the player’s knowledge on the LOs. It conducts the AAR sessions based on tutoring tactics and the natural language generation knowledge base. While discussing the actions performed, the students are also provided with the actual gameplay session’s videos to help them remind how they made the choices.

Evaluation Matrices and concluding remarks:

The lack of a definite notion of correctness makes it very hard for the designers to evaluate such systems. Instead of formalizing the evolution like other machine learning tasks, a set of Situational Judgment Tests(SJT) are administered to the students before and after exposure to the tutoring system. These SJTs are set of some case-study evaluation questions from which a user’s pattern of judgement is extracted. Then those patterns are then compared to the patterns of the domain experts in negotiations and found out how aligned the student’s pattern is to theirs. The authors describe the scenarios prepared by The U.S. Army Research Institute and ICT for bilateral negotiation in an Iraqi cultural context with examples from actual tests and corroborated with results. Though the researchers noticed significant increase in the SJT scores after these experiments, they couldn’t conclude that the improvements are solely because of the BiLAT system’s tutoring efficacy. Because the increase in judgmental ability could very well be because of the overall experience of the student during the gameplay. They show a comparison of the results depending on whether participants were commissioned or noncommissioned, and whether participants had prior negotiation experience in a non-western culture or not. Apart from SJT scores, they also collected some gameplay logs from which they modeled some no. of objectives related performance matrices. The author concludes the evaluation discussions saying that none of the above matrices were adequate enough to capture the actual performance of the models. Finally, they give some very insightful future direction on the BiLAT system, among which  the advanced use of NLP techniques for student modeling is a very interesting one. Overall, we believe that the paper is a very effective read and helps the novice learners of Artificial Negotiation domain to understand the basics of the field.

NB: The article is a critique on Kim, J. M., Hill Jr, R. W., Durlach, P. J., Lane, H. C., Forbell, E., Core, M., ... & Hart, J. (2009). BiLAT: A game-based environment for practicing negotiation in a cultural context. International Journal of Artificial Intelligence in Education, 19(3), 289-308.