In the simulation of human crowd behavior including evacuation planning, transportation management, and safety engineering in architecture design, the development of pedestrian model for higher behavior fidelity is an important task. To construct plausible facsimiles of real crowd movements, simulations should exhibit human behaviors for navigation, pedestrian decision-making, and social behaviors such as grouping and crowding. The research field is quite mature in some sense, with a large number of approaches that have been proposed to path finding, collision avoidance, and visually pleasing steering behaviors of virtual humans. However, there is still a clear disparity between the variety of approaches and the quality of crowd behaviors in simulations. Many social science field studies inform us that crowds are typically composed of multiple social groups (James, 1953; Coleman and James, 1961; Aveni, 1977). These observations indicate that one component of the complexity of crowd dynamics emerges from the presence of various patterns of social interactions within small groups that make up the crowd. Hence, realism in a crowd simulation may be enhanced when virtual characters are organized in multiple social groups, and exhibit human-like coordination behaviors. Motivated by the need for modeling groups in a crowd, we present a multi-agent model for large crowd simulations that incorporates socially plausible group behaviors. A computational model for multi-agent coordination and interaction informed by well- established Common Ground theory (Clark, 1996; Clark and Brennan, 1991) is proposed. In our approach, the task of navigation in a group is viewed as performing a joint activity which requires maintaining a state of common ground among group members regarding walking strategies and route choices. That is, group members communicate with, and adapt their behaviors to each other in order to maintain group cohesiveness while walking. In the course of interaction, an agent may present gestures or other behavioral cues according to its communicative purpose. It also considers the spatiotemporal conditions of the agent-group's environment in which the agent interacts when selecting a kind of motions. With the incorporation of our agent model, we provide a unified framework for crowd simulation and animation which accommodates high-level socially-aware behavioral realism of animated characters. The communicative purpose and motion selection of agents are consistently carried through from simulation to animation, and a resulted sequence of animated character behaviors forms not merely a chain of reactive or random gestures but a socially meaningful interactions. We conducted several experiments in order to investigate the impact of our social group interaction model in crowd simulation and animation. By showing that group communicative behaviors have a substantial influence on the overall distribution of a crowd, we demonstrate the importance of incorporating a model of social group interaction into multi-agent simulations of large crowd behaviors. With a series of perceptual user studies, we show that our model produces more believable behaviors of animated characters from the viewpoint of human observers. / Ph. D.

Twitter has become a popular data source in the recent decade and garnered a significant amount of attention as a surrogate data source for many important forecasting problems. Strong correlations have been observed between Twitter indicators and real-world trends spanning elections, stock markets, book sales, and flu outbreaks. A key ingredient to all methods that use Twitter for forecasting is to agree on a domain-specific vocabulary to track the pertinent tweets, which is typically provided by subject matter experts (SMEs). The language used in Twitter drastically differs from other forms of online discourse, such as news articles and blogs. It constantly evolves over time as users adopt popular hashtags to express their opinions. Thus, the vocabulary used by forecasting algorithms needs to be dynamic in nature and should capture emerging trends over time. This thesis proposes a novel unsupervised learning algorithm that builds a dynamic vocabulary using Probabilistic Soft Logic (PSL), a framework for probabilistic reasoning over relational domains. Using eight presidential elections from Latin America, we show how our query expansion methodology improves the performance of traditional election forecasting algorithms. Through this approach we demonstrate how we can achieve close to a two-fold increase in the number of tweets retrieved for predictions and a 36.90% reduction in prediction error. / Master of Science

Pedestrian crowd simulation explores crowd behaviors in virtual environments. It is extensively studied in many areas, such as safety and civil engineering, transportation, social science, entertainment industry and so on. As a common phenomenon in pedestrian crowds, grouping can play important roles in crowd behaviors. To achieve more realistic simulations, it is important to support group modeling in crowd behaviors. Nevertheless, group modeling is still an open and challenging problem. The influence of groups on the dynamics of crowd movement has not been incorporated into most existing crowd models because of the complexity nature of social groups. This research develops a framework for group modeling in agent-based pedestrian crowd simulations. The framework includes multiple layers that support a systematic approach for modeling social groups in pedestrian crowd simulations. These layers include a simulation engine layer that provides efficient simulation engines to simulate the crowd model; a behavior-based agent modeling layers that supports developing agent models using the developed BehaviorSim simulation software; a group modeling layer that provides a well-defined way to model inter-group relationships and intra-group connections among pedestrian agents in a crowd; and finally a context modeling layer that allows users to incorporate various social and psychological models into the study of social groups in pedestrian crowd. Each layer utilizes the layer below it to fulfill its functionality, and together these layers provide an integrated framework for supporting group modeling in pedestrian crowd simulations. To our knowledge this work is the first one to focus on a systematic group modeling approach for pedestrian crowd simulations. This systematic modeling approach allows users to create social group simulation models in a well-defined way for studying the effect of social and psychological factors on crowd’s grouping behavior. To demonstrate the capability of the group modeling framework, we developed an application of dynamic grouping for pedestrian crowd simulations.

Development of the Jordan Performance Appraisal System (JPAS) was completed in 1996. This study examined the factor structure of the classroom observation instrument used in the JPAS. Using observed classroom instructional quality ratings of 1220 elementary teachers of Grades 1-6 in the Jordan School District, this study estimated the factor structure of the data and the rater effect on relevant structural parameters. This study also tested for measurement invariance at the within and between levels across teachers of two grade-level groups (a) lower grades: Grades 1-3 and (b) upper grades: Grades 4-6. Factor structure was estimated using complex exploratory factor analysis (EFA) conducted on a subset of the original data. The analysis provided evidence of a three-factor model for the combined groups. The results of multiple confirmatory factor analyses (CFA) conducted using a different subset of the data cross-validated EFA results. Results from multilevel confirmatory factor analysis (MCFA) indicated the three-factor model fit best at both the within and the between levels, and that the intraclass correlation (ICC) was high (.699), indicating significant rater-level variance. Results from a multilevel multigroup confirmatory factor analysis (MLMG-CFA) indicated that the ICC was not significantly different between groups. Results also indicated configural, metric (weak factorial), and scalar (strong factorial) equivalence between groups. This study provided one of the first examples of how to estimate the impact of cluster-level variables such as rater on grouping variables nested at the within level. It provided an example of how to conduct a multilevel multigroup analysis on count data. It also disproved the assumption that counting classroom teaching behaviors was less subjective than using a categorical rating scale. These results will provide substantial information for future developments made to the classroom observation instrument used in the JPAS.