MAS Lab Projects

Current Projects


Organizational design, performance and adaptation

about We develop KB-ORG: a fully automated, knowledge-basedorganization designer for multi-agent systems. KB-ORG uses both application-level and coordination-level organization design knowledge to explore the combinatorial search space of candidate organizations selectively. We also show that KB-ORG designs effective, yet substantially different, organizations when given different requirements and environmental expectations.
people Mark Sims, Dan Corkill, Victor Lesser

Complex negotiation models in electronic commerce

about In electronic commerce markets where selfish agents behave individually, agents often have incomplete information as well as many constraints like budget, deadline. We are developing negotiation models in which agents often have to acquire multiple resources in order to accomplish a high level task with each resource acquisition requiring a separate negotiation thread. We take a set of elements into account, for example, deadline, outside option, market competition, multiple resources, and decommitment.
people Bo An, Victor Lesser

Approximately solving sequential games

about Games of incomplete information are notorious for their difficulty which usually makes finding an exact solution intractable. The problem is even harder when the game has multiple stages and sequential decision-making is required. We developed an anytime search algorithm for calculating approximate Bayes-Nash equilibria in sequential games of incomplete information. Experimental results demonstrate our algorithm's attractive anytime behavior which allows it to find good-enough solutions to large games within reasonable amounts of time.
people Hala Mostafa, Victor Lesser

Distributed Bayesian Networks

about In environments where observations and reasoning are distributed, gathering information consumes considerable resources. Moreover, having smaller networks reduces reasoning complexity. Instead of breaking the large or distributed network at random convenient points, understanding the relationships between nodes by identifying necessary evidences for reasoning helps reduce the size of the network. Context-specific independence provides grounds to identify a set of observations necessary for accurate reasoning that is smaller than the network structure itself.
people Yoonheui Kim, Victor Lesser

Organizationally motivated network routing

about Multi-Agent organization knowledge is traditionally not utilized by general-purpose wireless network routing algorithms normally used to support agent communication. We show that incorporating organization knowledge (otherwise available only to the application layer) in the network-layer routing algorithm increases bandwidth available at the application layer.
people Huzaifa Zafar, Dan Corkill, Victor Lesser

Learning distributed task allocation policies

about In the multi-agent setting, agents are simultaneously learning their interaction policies, which can result in low convergence and even divergence of the learning. We investigate the issue of scalable efficient Multi-Agent Reinforcement Learning (MARL) techniques by introducing a hierarchical organization with multi-leveled distributed supervision, and propose a general approach to integrating this supervision mechanism with MARLs to coordinate learning among agents. Each level in the organization structure is an overlay network.
people Chongjie Zhang, Victor Lesser

GILA: Integrated Learning

about We study ways to coordinate multiple learning agents, each with a different learning algorithm, and integrate their hypothesis to cooperatively and incrementally solve complex problems. We also investigate learning algorithms that can improve such coordination and cooperation. Air traffic flight planning is the primary application of this research project.
people Chongjie Zhang, Hala Mostafa, Dan Corkill, Victor Lesser

CNAS - Collaborative Network for Atmospheric Sensing

about CNAS is an experimental, agent based, power-aware sensor network for ground-level atmospheric monitoring. Due to characteristics such as sparse deployment of nodes, replenishable energy sources, and a need for turning off WiFi radios for most of the time, CNAS provides an interesting application for research in organizationally motivated routing, developing solar energy harvesting models, and multi-agent systems deployment.
people Dan Corkill, Huzaifa Zafar
papers [AAMAS-08], [ATSN08(Zafar,Corkill)], [ATSN08(Corkill)], [ATSN07]

Research Sponsors


Past Projects

about Integrating contributions received from other agents is an essential activity in multi-agent systems (MASs). Not only must related contributions be integrated together, but the confidence in each integrated contribution must be determined. We look specifically at the issue of confidence determination and its effect on developing "principled" highly collaborating MASs.

Fusing contributions from multiple agents

about Integrating contributions received from other agents is an essential activity in multi-agent systems (MASs). Not only must related contributions be integrated together, but the confidence in each integrated contribution must be determined. We look specifically at the issue of confidence determination and its effect on developing "principled" highly collaborating MASs.

Autonomous Negotiating Teams

about The Autonomous Negotiating Teams (ANTS) project seeks to address the problem of coordinating over constrained resources in an uncertain, real-time domain. In this domain, there are a number of radar based sensors that must track targets moving through the environment. The sensors must coordinate their activity to "triangulate" the exact position of the targets. Using a variety of techniques, which has pushed our technology to the next level, we have developed a set of negotiation strategies that operate on different levels of abstraction based on the immediate demands of the environment to coordinate the use of the sensors. Future work in this domain includes the development of methods for organizing large-scale collections of agents.

Explaining the Behavior of Multi-Agent Systems

about As multi-agent systems grow in scale and complexity it is increasingly difficult to understand and characterize the behavior of the system as a whole as well as individual agents. In order to help investigators understand these complex systems, we are developing analysis techniques and software tools that are founded on analytic approaches developed within the intelligence analysis and social network analysis communities. Specifically, we are using relational data representations that we believe can capture many of the important aspects of the behavior of multi-agent systems, and potentially capture far more of those behaviors than traditional data representations. These approaches are almost unknown in computer science, but we believe that they are uniquely suited to help understand the behavior of multi-agent systems.

Bounded Information Gathering

about The BIG (resource-Bounded Information Gathering) project is an information gathering agent that helps Internet shoppers cope with the proliferation of electronically available information. The process begins when a user enters in a desired product's characteristics, and search criteria. BIG then uses this information to perform the actual search and recommendation process, which includes generating and scheduling a plan of activities, searching through natural language reviews and descriptions, sophisticated text processing, reasoning about resource trade-offs, and the extraction and consolidation of relevant information. Once completed, BIG presents the user with the list of product descriptions which best satisfy the user's criteria.

Soft Real-Time Architecture

about Real-time control has become increasingly important as technologies are moved from the lab into real world situations or physical simulations. The complexity associated with these systems increases as control and autonomy are distributed, due to such issues as precedence constraints, shared resources, and the lack of a complete and consistent world view. The Soft Real-Time Architecture addresses these issues by providing a robust scheduling and execution subsystem capable of quantitatively reasoning over deadlines and resource constraints. This provides a useful layer of abstraction, enabling the agent's higher level reasoning components to operate at a more tractable level of granularity, without sacrificing fine-grained control and reactivity.