Equilibrium or near-equilibrium solutions to very large extensive form games are often computed by using abstractions to reduce the game size. A common abstraction technique for games with a large number of available actions is to restrict the number of legal actions in every state. This method has been used to discover equilibrium solutions for the game of no-limit heads-up Texas Hold'em. When using a solution to an abstracted game to play one side in the un-abstracted (real) game, the real opponent actions may not correspond to actions in the abstracted game. The most popular method for handling this situation is to translate opponent actions in the real game to the closest legal actions in the abstracted game. We show that this approach can result in a very exploitable player and propose an alternative solution. We use probabilistic mapping to translate a real action into a probability distribution over actions, whose weights are determined by a similarity metric. We show that this approach significantly reduces the exploitability when using an abstract solution in the real game.

The problem of environment design considers a setting in which an interested party aims to influence an agent's decisions by making limited changes to the agent's environment.  Zhang and Parkes [2008] first introduced the environment design concept for a specific problem in the Markov Decision Process setting. In this paper, we present a general framework for the formulation and solution of environment design problems. We consider both the case in which the agent's local decision model is known and partially unknown to the interested party, and illustrate the framework and results on a linear programming setting.  For the latter problem, we formulate an active, indirect elicitation method and provide conditions for convergence and logarithmic convergence. We relate to the problem of inverse optimization and also offer a game-theoretic interpretation of our methods.

We introduce a framework so that communities can exchange reputation information about agents in environments where agents are migrating between communities. We view the acquisition of the reputation information as a purchase and focus on the design of a payment function to facilitate the payment for information in a way that motivates communities to truthfully report reputation information for agents. We prove that in our proposed framework, honesty is the optimal policy and demonstrate the value of using a payment-function approach for the exchange of reputation information about agents between communities in multiagent environments. Using our payment function, each community is strengthened: it is able to reason more effectively about which agents to accept and can enjoy agents that are motivated to contribute strongly to the benefit of the community.

We propose a novel method for helping humans make good decisions in complex games, for which common equilibrium solutions may be too difficult to compute or not relevant. Our method leverages and augments humans' natural use of arguments in the decision making process. We believe that, if computers were capable of generating similar arguments from the mathematical description of a game, and presented those to a human decision maker, the synergies would result in better performance overall. The theory of reasoning patterns naturally lends itself to such a use. We use reasoning patterns to derive localized evaluation functions for each decision in a game, then present their output to humans. We have implemented this approach in a repeated principal-agent game, and used it to generate advice given to subjects. Experimental results show that humans who received advice performed better than those who did not.

In this paper, we study distributed algorithms for cooperative agents that  allow them to exchange their assigned tasks in order to reduce their team  cost. We define a new type of contract, called K-swaps, that describes multiple task exchanges among multiple agents at a time, which generalizes the concept of single task exchanges. We design a distributed algorithm that constructs all possible K-swaps that reduce the team cost of a given task allocation and show that each agent typically only needs to communicate a small part of its local computation results to the other agents. We then demonstrate empirically that K-swaps can reduce the team costs of several existing task-allocation algorithms significantly even if K is small.

Planning under uncertainty for multiple agents has flourished with the development of formal models such as multi-agent MDPs and decentralized MDPs. Despite their richness, applicability of these models remains limited due to their computational complexity. We present the class of event-detecting Multi-agent MDPs (eMMDPs), designed to detect multiple mobile targets by a team of sensor agents. We show that eMMDPs are NP-Hard and present a scalable 2-approximation algorithm for solving them using matroid theory and constraint optimization. Its complexity is linear in the state-space and number of agents, quadratic in the horizon, and exponential only in a small parameter that depends on the interaction among the agents. Despite the worst-case approximation ratio of 2, experimental results show that the algorithm produces nearoptimal policies for a range of test problems.

In this paper, we introduce an on-line, decentralised coordination algorithm for monitoring and predicting the state of spatial phenomena by a team of mobile sensors. These sensors have their application domain in disaster response, where strict time constraints prohibit path planning in advance. The algorithm enables sensors to coordinate their movements with their direct neighbours to maximise the collective information gain, while predicting measurements at unobserved locations using a Gaussian process. It builds upon the max-sum message passing algorithm for decentralised coordination, for which we present two new generic pruning techniques that result in speed-up of up to 92% for 5 sensors. We empirically evaluate our algorithm against several on-line adaptive coordination mechanisms, and report a reduction in root mean squared error up to 50% compared to a greedy strategy.

Strategyproof classification deals with a setting where a decision-maker must classify a set of input points with binary labels, while minimizing the expected error. The labels of the input points are reported by self-interested agents, who might lie in order to obtain a classifier that more closely matches their own labels, thus creating a bias in the data; this motivates the design of truthful mechanisms that discourage false reports. Previous work [Meir et al., 2008] investigated both decision-theoretic and learning-theoretic variations of the setting, but only considered classifiers that belong to a degenerate class. In this paper we assume that the agents are interested in a shared set of input points. We show that this plausible assumption leads to powerful results. In particular, we demonstrate that variations of a truthful random dictator mechanism can guarantee approximately optimal outcomes with respect to any class of classifiers.

This paper considers the execution of a Multi-Agent Plan in a partially observable environment, and faces the problem of recovering from action failures.  The paper formalizes a local plan repair strategy, where each agent in the system is responsible for controlling (monitoring and diagnosing) the actions it executes, and for autonomously repairing its own plan when an action failure is detected.  The paper describes also how to mitigate the impact of an action failure on the plans of other agents when the local recovery strategy fails.

We introduce planning games, a study of interactions of self-motivated agents in automated planning settings. Planning games extend STRIPS-like models of single-agent planning to systems of multiple self-interested agents, providing a rich class of structured games that capture subtle forms of local interactions. We consider two basic models of planning games and adapt game-theoretic solution concepts to these models.  In both models, agents may need to cooperate in order to achieve their goals, but are assumed to do so only in order to increase their net benefit. For each model we study the computational problem of finding a stable solution and provide efficient algorithms for systems exhibiting acyclic interaction structure.