Research > Multiagent Systems

Grant Number: W911NF-09-1-0464

ARMY RESEARCH OFFICE (ARO)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

Co-Principal Investigator: Adam Goodie, Psychology Department, UGA

Uninhabited agents such as UAVs are assuming increasingly prominent roles in contemporary wartime theaters. As greater numbers of UAVs are deployed, there is a need to reduce its reliance on human operators and transfer greater autonomy in decision making. Because UAVs may coexist with humans and human-controlled agents, computational models of human decision making are needed. Three empirical studies will measure and validate probability assessment in tasks that simulate those in operating theaters, and will provide data for formulating these models. The studies will establish the validity of and help model the probability judgment of humans engaged in an adversarial role expected in UAV operating theaters, and will provide critical data for formulating the behavioral models of significant role players in a UAV's operating theater. Additionally, this extends the research of human-computer interaction to a UAV domain, and it promises robust solutions to the complex problem of accounting for both an operator's cognitive biases and the biases of an adversary.

Project Funding: $314472

Project Period: 2009 - 2012

Grant Number: IIS-1444182

NATIONAL SCIENCE FOUNDATION (NSF)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

This new, catalytic U.S.-Netherlands research collaboration addresses renewable energy-driven smart grids. Renewable energy sources include resources that are regularly replenished, such as sunlight, wind, rain, tidal waves, and geothermal heat. To pursue innovative approaches for managing the uncertainty of renewable energy sources, the U.S. principal investigator (PI) and a graduate student will visit the Netherlands to begin a collaboration with counterparts at the Delft University of Technology, a leader in European smart energy research. There they intend to work together to improve current smart grid technology for better prediction of consumer demand in the face of uncertain power generation, as is often the case in renewable energy systems. If successful, their preliminary results should contribute to improving bidirectional communication between grid operators and consumers. Early results and follow-on research may have broader impact by shaping management strategies through new approaches to modeling consumer energy usage. Success could mean better long-term prediction by employing new artificial intelligence approaches (AI), i.e., smart controls for power grids. The team expects to identify the challenges posed by the uncertainty of renewable energy generation and begin investigating intelligent methods for meeting these challenges in two priority areas: (a) planning for decentralized power generation and storage, and (b) managing congestion in grids due to asynchrony between renewable energy supply and consumer demand. The PI will work with an experienced team of eminent Dutch researchers in AI, power systems, and technology policy. They will have real operating and energy-use data from a medium voltage grid in Netherlands and intend to start developing scalable algorithms for individual decision making in multi-agent settings. Further, broader impacts are anticipated from this collaboration with an introduction of smart energy systems into research and teaching at the University of Georgia, thereby contributing to training U.S. undergraduate and graduate students in an innovative and rapidly growing energy sector with industrial relevance.

Project Funding: $33608

Project Period: 2015 - 2016

Grant Number: FA9550-08-1-0429

AIR FORCE OFFICE OF SCIENTIFIC RESEARCH (AFOSR)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

Co-Principal Investigator: Adam Goodie, Psychology Department, UGA

Uninhabited agents such as UAVs are increasingly being deployed in wartime theaters in both reconnaissance and active engagement roles. With the objective of reducing reliance on their human controllers, potential decision-making technology for these agents must choose optimally among several actions in a timely manner while operating in an uncertain environment that could be populated by both human and robotic agents. Within this setting, the proposed research will focus on the decision-making process of an agent that acts to maximize its preferences in an uncertain environment shared with other agents having similar or conflicting objectives. This problem of individual decision-making in uncertain multi-agent settings is formalized using a new theory that combines decision theory as formalized by partially observable Markov decision processes (POMDPs) with elements of Bayesian games and interactive epistemology. The aims of this research are to investigate ways to model the uncertainty in the strategic reasoning of agents, understand the sources of computational complexity within the theory and identify ways to mitigate their impact without significant losses in the optimality of the decision-maker. This will allow a general and scalable framework for individual decision-making in uncertain and large-scale settings, characterized by large spaces of physical states, actions, observations, and a large number of agents. Because the agent may share its environment with humans and human-controlled agents, computational models of human decision-making are needed to anticipate the behavior of these agents. While several prominent psychological models of decision-making exist, these are not widely agreed upon and lack aspects such as strategic decision making that are central to multi-agent settings. With the objective of developing empirically informed models of human decision-making, this research will focus on modeling strategic behavioral data related to the theory of mind (TOM). New empirical studies that enhance the understanding of strategic human behavior and contribute to TOM are proposed.

Project Funding: $237990

Project Period: 2008 - 2010

Grant Number: IIS-0845036

NATIONAL SCIENCE FOUNDATION (NSF)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

Research under this award is developing efficient and effective methods for strategic decision making by an individual artificial agent cohabiting with other agents in uncertain environments. For example, how should an autonomous unmanned aerial vehicle decide between closer surveillance of a possible fugitive or intercepting the target who may be aware of the monitoring? Toward this goal, the research is identifying the sources of computational complexity and understanding the conflicting interrelationship between computational efficiency and decision-making effectiveness. This problem of individual decision making in uncertain multiagent settings is formalized using a recognized framework that combines the decision-theoretic paradigm of partially observable Markov decision processes (POMDPs) with elements of Bayesian games and interactive epistemology. In this framework, called interactive POMDP (I-POMDP), the research utilizes innovative ways of minimally modeling contextual knowledge in multiagent settings, exploits novel decision-making heuristics and embedded structure in problems. Integration of research and education is manifest in the development and delivery of a multi-disciplinary course on strategic decision making under uncertainty, which integrates and compares normative theories with real human decision- making behavior. By combining aspects of decision and game theories, both of which seek to understand normative ways of decision making, with attention to real human decision-making behavior, this research is contributing to long-term research and development of artificial agents that can assist with rational, long-term decision making and planning in areas including emergency response, environmental sustainability, autonomous vehicles and many others.

Project Funding: $429663

Project Period: 2009 - 2014

Grant Number: IIS-1346942

NATIONAL SCIENCE FOUNDATION (NSF)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

Co-Principal Investigator: Makoto Yokoo, Kyushu University, Japan

The overall goal of this research is to understand new characterizations of equilibria for games that are realistic and complex, and identify scalable ways of computing the equilibria. In keeping with this goal, the objective of this EAGER proposal is to investigate algorithms for computing strategy profiles in exact and approximate equilibrium in the context of repeated games with imperfect private monitoring. This general class of games exhibits a wide swath of practical applications that include repeated first-price auctions with private outcomes, understanding secret price wars between firms and analyzing extended partnership models. Until recently, the structure of equilibria in these games was very poorly understood. A complete mathematical characterization of equilibria for infinitely repeated games with private monitoring developed recently introduces the concept of a finite state equilibrium (FSE). This correlated sequential equilibrium is defined between strategies that manifest as finite state automata, and its verification involves the formulation and solution of a partially observable Markov decision process (POMDP). Preliminary research demonstrates the feasibility of utilizing POMDPs to verify symmetric FSE in simple repeated games. The proposed research will take the leap from verifying FSE to developing algorithms for computing strategy profiles in equilibrium for the games. A key consideration would be computational scalability of the algorithms as games with an increasing number of players and dimensions are considered, thereby laying the foundation for practical applications. Consequently, this research conceptualizes approximate FSE for such repeated games. It will design novel algorithms for computing ǫ-approximate FSE that utilize error-bounded approximate solutions of POMDPs. The high payoff from identifying and understanding equilibrium behavior in the sophisticated interactions modeled by the studied class of games makes this research well suited to the EAGER mechanism, while its early nature does not make it a good fit for submission to a regular program until supportive results are available.

Project Funding: $150153

Project Period: 2013 - 2015

Grant Number: N000141310870

OFFICE OF NAVAL RESEARCH (ONR)

Principal Investigator: Prashant Doshi, Computer Science Department, UGA

A wide swath of multidisciplinary areas in autonomous multiagent systems benefit from modeling interacting agents. However, the space of possible mental models is generally very large and grows disproportionately as the interaction progresses. Sometimes, application constraints and data may limit this space. The objective of the proposed research is to compress the model spaces in principled and domain-independent ways. In keeping with this objective, my approach is to partition spaces by forming equivalence classes of models and retaining a single representative from each class. I propose innovative methods for compression that are both exact as these do not result in any associated loss of information for the modeling agent (lossless) but which are computationally intensive, and approximate that are computationally more efficient but lossy. The broad usefulness of the techniques will be established by evaluating the impact of the compression on the scalability and quality of planning in large and partially observable multiagent settings, of Bayesian plan recognition with an exhaustive library of plans, and of equilibrium computation in incomplete information games exhibiting large type spaces. An additional objective of the experimentation will be to inform easy-to-use heuristics that could guide the approximate compression of the model spaces concomitantly yielding acceptable error. As naval decision making and planning often occurs in the context of imperfect information about the adversary, minimizing the large model spaces while bounding approximation error will facilitate strategic analyses by future naval decision-support systems that is comprehensive, situated and justifiable. The diverse applicability of the research is indicative of its high payoff nature but it is accompanied by the risk of whether the model spaces are generally compressible in principled ways to a degree that significantly improves scalability. Results from my preliminary research serve to mitigate this risk.

Project Funding: $570327

Project Period: 2013 - 2016