Gita Sukthankar

Research directions for service-oriented multiagent systems

Todays service-oriented systems realize many ideas from the research conducted a decade or so ago in multiagent systems. Because these two fields are so deeply connected, further advances in multiagent systems could feed into tomorrows successful service-oriented computing approaches. This article describes a 15-year roadmap for service-oriented multiagent system research.

Dynamic shadow elimination for multi-projector displays

A major problem with interactive displays based on front-projection is that users cast undesirable shadows on the display surface. This situation is only partially addressed by mounting a single projector at an extreme angle and pre-warping the projected image to undo keystoning distortions. This paper demonstrates that shadows can be muted by redundantly illuminating the display surface using multiple projectors, all mounted at different locations. However, this technique alone does not eliminate shadows: multiple projectors create multiple dark regions on the surface (penumbral occlusions). We solve the problem by using cameras to automatically identify occlusions as they occur and dynamically adjust each projectors output so that additional light is projected onto each partially-occluded patch. The system is self-calibrating: relevant homographies relating projectors, cameras and the display surface are recovered by observing the distortions induced in projected calibration patterns. The resulting redundantly-projected display retains the high image quality of a single-projector system while dynamically correcting for all penumbral occlusions. Our initial two-projector implementation operates at 3 Hz.

Shadow elimination and occluder light suppression for multi-projector displays

Two related problems of front projection displays, which occur when users obscure a projector, are: (i) undesirable shadows cast on the display by the users, and (ii) projected light falling on and distracting the users. This paper provides a computational framework for solving these two problems based on multiple overlapping projectors and cameras. The overlapping projectors are automatically aligned to display the same dekeystoned image. The system detects when and where shadows are cast by occluders and is able to determine the pixels, which are occluded in different projectors. Through a feedback control loop, the contributions of unoccluded pixels from other projectors are boosted in the shadowed regions, thereby eliminating the shadows. In addition, pixels, which are being occluded, are blanked, thereby preventing the projected light from falling on a user when they occlude the display. This can be accomplished even when the occluders are not visible to the camera. The paper presents results from a number of experiments demonstrating that the system converges rapidly with low steady-state errors.

Robust recognition of physical team behaviors using spatio-temporal models

This paper presents a framework for robustly recognizing physical team behaviors by exploiting spatio-temporal patterns. Agent team behaviors in athletic and military domains typically exhibit an observable structure characterized by the relative positions of teammates and external landmarks, such as a team of soldiers ambushing an opponent or a soccer player moving to receive a pass. We demonstrate how complex team relationships that are not easily expressed by region-based heuristics can be modeled from data and domain knowledge in a way that is robust to noise and spatial variation. To represent team behaviors in our domain of MOUT (Military Operations in Urban Terrain) planning, we employ two classes of spatial models: 1) team templates that encode static relationships between team members and external landmarks; and 2) spatially-invariant Hidden Markov Models (HMMs) to represent evolving agent team configurations over time. These two classes of models can be combined to improve recognition accuracy, particularly for behaviors that appear similar in static snapshots. We evaluate our modeling techniques on large urban maps and position traces of two-person human teams performing MOUT behaviors in a customized version of Unreal Tournament (a commercially available first-person shooter game).

A cost minimization approach to human behavior recognition

This paper presents a cost minimization approach to the problem of human behavior recognition. Using full-body motion capture data acquired from human subjects, our system recognizes the behaviors that a human subject is performing from a set of military maneuvers, based on the subjects motion type and proximity to landmarks. Low-level motion classification is performed using support vector machines (SVMs) and a hidden Markov Model (HMM); output from the classifier is used as an input feature for the behavior recognizer. Given the dynamic and highly reactive nature of the domain, our system must handle behavior sequences that are frequently interrupted and often interleaved. To recognize such behavior sequences, we employ dynamic programming in conjunction with a behavior transition cost function to efficiently select the most parsimonious explanation for the humans actions. We demonstrate that our system is robust to action classification errors and deviations by the human subject from the expected set of behaviors. Our approach is well suited for incorporation into synthetic agents that cooperate or compete against human subjects in virtual reality training environments.

Incremental Relabeling for Active Learning with Noisy Crowdsourced Annotations

Crowd sourcing has become an popular approach for annotating the large quantities of data required to train machine learning algorithms. However, obtaining labels in this manner poses two important challenges. First, naively labeling all of the data can be prohibitively expensive. Second, a significant fraction of the annotations can be incorrect due to carelessness or limited domain expertise of crowd sourced workers. Active learning provides a natural formulation to address the former issue by affordably selecting an appropriate subset of instances to label. Unfortunately, most active learning strategies are myopic and sensitive to label noise, which leads to poorly trained classifiers. We propose an active learning method that is specifically designed to be robust to such noise. We present an application of our technique in the domain of activity recognition for eldercare and validate the proposed approach using both simulated and real-world experiments using Amazon Mechanical Turk.

Community detection in dynamic social networks: a game-theoretic approach

Most real-world social networks are inherently dynamic and composed of communities that are constantly changing in membership. As a result, recent years have witnessed increased attention toward the challenging problem of detecting evolving communities. This paper presents a game-theoretic approach for community detection in dynamic social networks in which each node is treated as a rational agent who periodically chooses from a set of predefined actions in order to maximize its utility function. The community structure of a snapshot emerges after the game reaches Nash equilibrium; the partitions and agent information are then transferred to the next snapshot. An evaluation of our method on two real world dynamic datasets (AS-Internet Routers Graph and AS-Oregon Graph) demonstrates that we are able to report more stable and accurate communities over time compared to the benchmark methods.

Motif Discovery and Feature Selection for CRF-based Activity Recognition

Due to their ability to model sequential data without making unnecessary independence assumptions, conditional random fields (CRFs) have become an increasingly popular discriminative model for human activity recognition. However, how to represent signal sensor data to achieve the best classification performance within a CRF model is not obvious. This paper presents a framework for extracting motif features for CRF-based classification of IMU (inertial measurement unit) data. To do this, we convert the signal data into a set of motifs, approximately repeated symbolic sub sequences, for each dimension of IMU data. These motifs leverage structure in the data and serve as the basis to generate a large candidate set of features from the multi-dimensional raw data. By measuring reductions in the conditional log-likelihood error of the training samples, we can select features and train a CRF classifier to recognize human activities. An evaluation of our classifier on the CMU Multi-Modal Activity Database reveals that it outperforms the CRF-classifier trained on the raw features as well as other standard classifiers used in prior work.

Modeling information diffusion and community membership using stochastic optimization

Communities are vehicles for efficiently disseminating news, rumors, and opinions in human social networks. Modeling information diffusion through a network can enable us to reach a superior functional understanding of the effect of network structures such as communities on information propagation. The intrinsic assumption is that form follows function-rational actors exercise social choice mechanisms to join communities that best serve their information needs. Particle Swarm Optimization (PSO) was originally designed to simulate aggregate social behavior; our proposed diffusion model, PSODM (Particle Swarm Optimization Diffusion Model) models information flow in a network by creating particle swarms for local network neighborhoods that optimize a continuous version of Hollands hyperplane-defined objective functions. In this paper, we show how our approach differs from prior modeling work in the area and demonstrate that it outperforms existing model-based community detection methods on several social network datasets.

Link prediction in multi-relational collaboration networks

Traditional link prediction techniques primarily focus on the effect of potential linkages on the local network neighborhood or the paths between nodes. In this paper, we study the problem of link prediction in networks where instances can simultaneously belong to multiple communities, engendering different types of collaborations. Links in these networks arise from heterogeneous causes, limiting the performance of predictors that treat all links homogeneously. To solve this problem, we introduce a new link prediction framework, Link Prediction using Social Features (LPSF), which weights the network using a similarity function based on features extracted from patterns of prominent interactions across the network.