MASTERS @Tandy School of Computer Science, TU

We are interested in developing negotiation frameworks using which self-interested agents can improve their payoffs in non-cooperative environments. We are studying a number of different approaches under varying assumptions about interaction history, knowledge requirements, and agent capabilities.

• Improving efficiencies of envy-free divisions:Researchers have developed procedures for dividing up goods between self-interested parties such that the allocation is envy-free, i.e., when every party (agent) believes that its share is not less than anyone else's share. These procedures, however, are not efficient (in the sense of pareto optimality) in general. Envy-free procedures allow agents to ignore the utility metrics of other agents if they are satisfied with a fare share of the goods being divided. We are interested in studying augmentations of these procedures in which agents use models of the decision strategies or utility metrics of other agents to try to obtain more than their fare share without sacrificing the envy-free guarantee.
• Bayes net based agent models: Modeling using Bayesian networks Relationship between agents in a society can be represented using a Bayesian network where the topology of the network together with the conditional and prior probabilities represent an agent's view of the influence of different factors on outcomes of agent interactions. Such a Bayes net model can also aid an agent in its negotiation with other agents. We are particularly interested in procedures to be used by an agent to create a favorable negotiation context in which negotiation takes place between two agents. We provide a decision mechanism by which an agent can take actions to create a favorable negotiation context in addition to choosing a negotiation offer that is likely to be accepted by the other agent. We are also interested in decision procedures that elicit useful information about other agents that flesh out the Bayes net model. We have developed a maximin entropy decision procedure that allows the modeling agent to choose actions to produce guaranteed minimal improvement of the model accuracy.
• Risk evaluation in partner selection: We consider situations where a rational agent has to choose one of several contenders to enter into a partnership. We assume that the agent has a model of the likelihood of different outcomes and corresponding utilities for each such partnership. Given a fixed, finite number of interactions, the problem is to choose a particular partner to interact with where the goal is to maximize the sum of utilities received from all the interactions. We develop a multinomial distribution based mechanism for partner selection and contrast its performance with other well-known approaches which provide exact solution to this problem for infinite interactions.
• Contracting in Supply Chains: We assume an open supply chain environment where manufacturers award sub-contracts to bidding suppliers using an auction scheme. Task scheduling decisions for a supplier should be targeted towards creating a schedule that is flexible in accommodating short-notice tasks or schedule adjustments due to unforeseen events. We believe that task scheduling heuristics will be crucial in determining the competitiveness of a supplier in the marketplace. This is particularly true when a supplier can provide scarce resources or services that other suppliers cannot provide and hence can significantly increase its contract prices and, therefore, profitability.
• Learning in repeated single-stage games: Multiagent learning literature has looked at iterated two-player games to develop mechanisms that allow agents to learn to converge on Nash Equilibrium strategy profiles. Often, in general sum games, a higher payoff can be obtained by both players if one chooses not to respond optimally to the other player. We have found that in certain situations, modeling agents, who select actions proportional to expected utility based on observed probability distributions of opponent actions, can perform better than those that converge to Nash Equilibrium. We also experiment with an interesting action revelation strategy that can give the revealer better payoff on convergence than a non-revealing approach.