• Home
  • User Menu
  • Calendar
  • Community
  • News
• People
  • Participant Students
  • Advisory Board
  • Affiliate Faculty
  • Project Staff
• Showcase
  • Research Projects
  • Publications
  • Promotional Materials
  • Other Products
• Credits
• Links
  • National Science Foundation
  • FIT
  • UCF
  • ECE Department at FIT
  • School of EECS at UCF
  • REU Program at NSF
  • Search for other REU sites at NSF
• FAQ
  • About AMALTHEA
  • For Prospective Students
• Contact

Research Projects

Each AMALTHEA REU team worked on a chosen research topic. Here you will find descriptions of these topics along with the posters that the teams presented during AMALTHEA's Symposium at the end of the summer experience. The posters are clickable links that will allow you to view them in better (higher) resolution. More details on the teams' research topics can be found in their technical reports (TRs; see the link below each project description). Finally, some of this research was published as conference and journal papers, which you can find under the Publications page.

	Title: Modeling Group Dynamics in Virtual Worlds By: Chris Usher Graduate mentor(s): Fahad Shah Faculty mentor(s): Dr. Gita Sukthankar Abstract: In this study, we examine human social interactions within virtual worlds and address the question of how group interactions are affected by the surrounding game environment. To investigate this problem, we analyzed conversational data collected from Second Life, a massively multi-player online environment that allows users to construct and inhabit their own 3D world. Our data collection chatbots were created to be sufficiently lifelike to be indistinguishable from other human participants to casual observers, so as not to perturb neighboring social interactions. Using our partitioning algorithm, we separated continuous chat logs from each region into separate conversations which were used to construct a social network of the participants. Based on statistical studies of social network measures, we conclude that there are significant regional differences between social networks formed in Second Life. TR: Usher, C., Shah, F., and Sukthankar, G. (2009) Modeling Group Dynamics in Virtual Worlds, Technical Report TR-2009-05, The AMALTHEA REU Program, Summer 2009 [PDF]
	Title: Kernel similarity scores for outlier detection in mixed-attribute data sets By: David Foregger, Julie Manuel Graduate mentor(s): Ruben Ramirez-Padron Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: The majority of current outlier detection methods are limited to data sets with all attributes of the same type. The few methods that deal with data sets with mixed numerical and categorical data are complex and explicitly depend on the data types of the attributes, which makes them difficult to extend to other types of mixed-attribute data. Kernel functions can be seen as similarity functions defined on any set of objects. Kernel functions on different feature representations can be combined by using a set of well established rules. To the best of our knowledge, however, current kernel-based outlier detection methods have been defined only on single-type data sets, and the kernel functions used have been components of relatively complex statistical techniques. This work proposes simple but effective kernel-based outlier detection methods for single type and mixed-attribute data sets. Three different kernel similarity scores are introduced in order to rank points for outlier detection. For arbitrary mixed-attribute data sets, different kernel and score combinations are used to obtain a single similarity score. It is shown that our approach outperforms the Greedy and AVF algorithms on categorical data sets and Otey et al's algorithm for data sets with numerical and categorical attributes. TR: Foregger, D., Manuel, J., Ramirez-Padron, R., and Georgiopoulos, M. (2009) Kernel similarity scores for outlier detection in mixed-attribute data sets, Technical Report TR-2009-04, The AMALTHEA REU Program, Summer 2009 [PDF]
	Title: Social Network Analysis for Target Recognition in Swarm Robotics By: Michael C. Koval, Adina E. S. Rubinoff Graduate mentor(s): Mahsa Maghami Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: Over the past few years, swarm-based systems have emerged as an attractive and promising approach for implementing distributed autonomous systems. This is useful in dierent applications, such as automatic target recognition (ATR). In ATR, the most pressing concern is the accuracy of the system in detecting and recognizing the targets. To fulfill this requirement, previous approaches require more than one agent to classify a target, which decreases the number of recognition errors. Increasing the number of agents required to recognize a task will make the system more accurate; however, it delays recognition, which is not acceptable in most applications. In addition, agents are never completely identical in the real world; even two identical agents will have sensors with dierent error tolerances. Unfortunately, previous approaches fail to distinguish accurate robots from less accurate ones. In this paper we propose a novel algorithm, named the NetBots algorithm, which improves accuracy in ATR by introducing a social network to the swarm system. This social network keeps track of the robots' target detection histories and is used to estimate the robots' accuracy. Any time two robots deciding a task agree with the majority decision, a link is formed between them. The PageRank algorithm is then used to rank the robots based on the number and ranks of the robots linking to them. Robots that are more accurate will agree with the majority more often, so they will have more links and therefore a higher rank. When deciding a task, higher ranked robots have more influence, so inaccurate robots' votes will not negatively impact the majority decision. This approach increases the overall accuracy of the system by 1.76 percent over 30 tasks, compared to other approaches with no social network. TR: Koval, M.C., Rubinoff, A.E.S., Maghami, M., and Georgiopoulos, M. (2009) Social Network Analysis for Target Recognition in Swarm Robotics, Technical Report TR-2009-03, The AMALTHEA REU Program, Summer 2009 [PDF]
	Title: Learning Approximate Isometries using Multi-Dimensional Scaling based on Radial Basis Function Neural Networks By: Ryan Gonet, RuiZhi Yu Graduate mentor(s): Mingbo Ma Faculty mentor(s): Dr. Georgios C. Anagnostopoulos Abstract: This paper presents a new multi-dimensional scaling (MDS) technique for generating approximate isometries by way of using Radial Basis Function neural networks (RBF-NN). Metric MDS using the Sammon mapping1 is available and commonly used to create approximate isometries in order to perform data visualization. However, classical-Sammon mapping has the disadvantage of being unable to interpolate or extrapolate novel samples. Other techniques, such as one using Multi-Layer Perceptron (MLP) to implement the mapping, cannot map datasets where only dissimilarities between patterns are known. By using RBF-NNs to train a model to map patterns from a higher dimension into a lower dimension, one can both perform dimensionality reduction on datasets consisting only of dissimilarities and also interpolate and extrapolate novel patterns. For training, both the delta rule, through conjugate gradient descent and the limited-memory Broyden-Fletcher-Goldfarb-Shanno quasi-Newton method were used. Step lengths were calculated using a line search that finds step lengths obeying the strong Wolfe conditions. Experimental results using a prototype implementation of RBF-NNs to perform Sammon mapping demonstrate that approximate isometries can be found. They also show that datasets consisting of only dissimilarities can still be visualized, where the reduction shows a structural relationship between patterns. This new ability, in conjunction with the ability to interpolate and extrapolate previously unseen patterns, creates novel opportunities for the application of MDS using Sammon mapping as a visualization technique. TR: Gonet, R., Yu, R., Ma, M., and Anagnostopoulos, G.C. (2009) Learning Approximate Isometries using Multi-Dimensional Scaling based on Radial Basis Function Neural Networks, Technical Report TR-2009-02, The AMALTHEA REU Program, Summer 2009 [PDF]
	Title: Designing ART-based Classifiers through Multi-Objective Memetic Evolution By: Timothy R. Mersch, Oriana X. Wen Graduate mentor(s): Rong Li Faculty mentor(s): Dr. Georgios C. Anagnostopoulos Abstract: In multi-objective optimization, the two pivotal attributes of the Pareto front are density and diversity. A high-density, high-diversity Pareto front is characterized by an ample set of significantly unique solutions. With respect to classifier models, obtaining a well-sampled Pareto front maximizes the ability to identify, along the trade-o curve relating model complexity and accuracy, the best-generalizing model on a testing set. Evolutionary algorithms operating in the multi-objective space have proven, to a certain extent, successful in optimizingART-based classifiers and approaching the Pareto-optimal set. Building on previous work, a memetic genetic algorithm has been devised that evolves in parallel subpopulations of Fuzzy ARTMAP classifier models in order to attain a Pareto front with improved density and diversity. The genetic algorithm works under a quantized objective space, wherein the population of individual models is subdivided by complexity. The incorporation of specialized structures, from the Hall of Fame to the gene pool, with systematic mechanisms, including a mutation selection scheme and simulated annealing, produces a memetic algorithm that takes advantage of global exploration and localized exploitation. Safeguards against duplication and the combined forces of global and local search manifest substantial enhancement of the density and diversity of the Pareto front. Two sets of experiments were conducted to optimize mutation and to evaluate this genetic algorithm against a simpler implementation. The first series of experiments favors a uniform selection of individuals for mutation. The second series of experiments reveals that the genetic algorithm improves upon a pre-existing multi-objective evolutionary algorithm that optimizes ART-based classifiers. The comparisons between the two algorithms were with either random or trained initialization and with dierent subpopulation sizes. Overall, the novel approach is shown to be superior in terms of hyper-area, density, and two-set coverage of the final Pareto front. Also, the champion networks produced by the novel approach exhibit greater generalization power over those of the previous work. TR: Mersch, T.R., Wen, O.X., Li, R., and Anagnostopoulos, G.C. (2009) Designing ART-based Classifiers through Multi-Objective Memetic Evolution, Technical Report TR-2009-01, The AMALTHEA REU Program, Summer 2009 [PDF]
	Title: Iterative Inner Solvers for Revised Simplex SVM Training By: Eric P. Astor, Winnie J. Lung Graduate mentor(s): Ruben Ramirez-Padron, Christopher G. Sentelle Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: Support Vector Machine (SVM) training is equivalent to solving a large constrained optimization problem. Much work has been spent on decompositional optimization methods for this problem, but non-decompositional approaches have only recently regained attention. Notably, Sentelle’s work in applying Rusin’s revised simplex method to SVM training demonstrated a significantly shorter training time for complex problems, but requires much more memory than the competing decompositional methods. We propose a revision to Sentelle’s approach, replacing his direct inner solver with an iterative solver: the preconditioned conjugate residual method. Results show an unexpected performance penalty, due to extreme ill-conditioning of the inner problems; avoiding this may require the creation of a specialized functional preconditioner. TR: Astor, E.P., Lung, W.J., Ruben Ramirez-Padron, Christopher G. Sentelle and Georgiopoulos, M. (2008) Iterative Inner Solvers for Revised Simplex SVM Training , Technical Report TR-2008-06, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Interactively Evolved Modular Neural Networks for Agent Control By: Jessica C. Sparks , Roberto Miguez Graduate mentor(s): John Reeder Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: As the realism in games continues to increase, through improvements in graphics and 3D engines, more focus is placed on the behavior of the simulated agents that inhabit the simulated worlds. The agents in modern video games must become more life like in order to seem to belong in the environments they are portrayed in. Many modern AI’s achieve a high level of realism but this is accomplished through significant developer time spent scripting the behaviors of the Non-Playable Characters or NPC’s. These agents will behave in a believable fashion in the scenarios they have been programmed for but do not have the ability to adapt to new situations. In this paper we introduce a modularized, real-time, co-evolution training technique to evolve adaptable agents with life like behaviors. Experiments conducted produced very promising results regarding efficiency of the technique, and demonstrate potential avenues for future research. TR: Sparks, J.C., Miguez, R., John Reeder, and Georgiopoulos, M. (2008) Interactively Evolved Modular Neural Networks for Agent Control, Technical Report TR-2008-05, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Detecting Outliers in Categorical Data Sets Using Non-Derivable Itemsets By: Michelle Fox , Gary Gramajo Graduate mentor(s): Anna Koufakou Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: Outlier Detection is a research field with many applications, such as detecting credit card fraud or network intrusions. Most previous research focused on numerical data and pair-wise distances among data points to detect outliers. Nevertheless, most categorical data sets lack straightforward mapping to numerical values and approaches that rely on computing distances do not apply so easily. Recently, a few outlier methods were proposed for categorical datasets using the concept of Frequent Itemsets (FIs). The number of generated FIs can be far too high, especially in the case of large, dense datasets, containing a high number of categorical values. There has been much research towards summarizing and/or condensing the FIs in a dataset to address this issue. However these ideas have not been applied directly to the field of outlier detection. In this report, we explore the effect of using a condensed representation of Frequent Itemsets, called Non-Derivable Itemsets (NDI), on the accuracy and efficiency of an outlier detection method. Our experimental results indicate that NDI-based Outlier Detection offers significant gains in terms of speed and scalability over a FI-based outlier detection, while maintaining comparable detection accuracy. TR: Fox, M.S., Gramajo, G.E., Koufakou, A. and Michael Georgiopoulos. (2008) Detecting Outliers in Categorical Data Sets Using Non-Derivable Itemsets , Technical Report TR-2008-04, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Development of a Large Vocabulary Continuous Speech Recognition System for Rich Transcription Evaluation Using HTK By: David A. Wax , Noah A. Larsen, Matthew J. Furstossc, and Veton Z. Kepuska Graduate mentor(s): Faculty mentor(s): Abstract: The focus of this research is to build a large vocabulary continuous speech recognition (LVCSR) system that converts speech to text in accordance with the National Institute of Standards and Technology(NIST) Rich Transcription (RTE) Evaluation requirements. The result of the current effort will serve as a baseline for future work in the development of an advanced speech recognition system based on WUW technology.1 Cambridge University's HTK Speech Recognition Toolkit Version 3.4 serves as the engine in this process. In order to create a sufficiently large speech dataset, multiple corpora are combined, including TIMIT, and NIST RTE 2006 (RT06) and 2007 (RT07) data. Recognition testing and evaluation is performed under a variety of different conditions to find the ideal parameters for optimum accuracy. Modifiable factors include insertion penalties (IP), language models, phonetic questioning, bootstrapping, and skip states. Performance is measured by word error rate (WER). The addition of insertion balancing consistently improved WER at both phone-level and word- level, while the removal of TIMIT shibboleth sentences demonstrated no significant change in WER. Phonetic questioning effectively improved computation time without a significant increase of WER. Training and testing on TIMIT corpus data with the implementation of a language model attained the lowest WER of 78.74%. Although 78.74% WER is higher than what other research has achieved with HTK,2 the addition of the language model improved WER by a relative difference of 48.97%. Additionally, performance is better than expected for using only 1 Gaussian Mixture and 3 output states per HMM. TR: Wax, D.A., Larsen, N.A., Furstoss, M.J., and Kepuska, V.Z. (2008) Development of a Large Vocabulary Continuous Speech Recognition System for Rich Transcription Evaluation Using HTK , Technical Report TR-2008-03, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Multi-stage Automatic License Plate Location & Recognition By: Rong Li , Musa Yassin Fort Graduate mentor(s): Faculty mentor(s): Dr. Georgios C. Anagnostopoulos Abstract: Since as early as 1970’s, the need of an Automatic License Plate Recognition system (ALPR) has arose based on the need to implement law enforcement and traffic control on transportation systems. This area has motivated a large amount of research effort, and various approaches and solutions have been proposed and implemented. Existing ALPR systems typically consist of modules addressing the following three tasks: license plate localization, character segmentation, and character recognition. In this paper, a novel ALPR system is proposed where a new de-skewing stage is added between license plate localization and character segmentation stages. This new stage allow the system to process images that are taken at an angle with respect to the LP’s normal and from a relatively close distance, which typically results in skew distortion. The additional stage ensures that the characters are all lined up in a horizontal line with the same height, which allows precise character segmentation; furthermore, it rectiffies the characters and allows simple recognition techniques such as cross-correlation to yield good classification results. In addition to the de-skewing stage, a license plate localization method is also proposed. It shares some common ground with the current approaches, as it is based on vertical edge density. Note that images with complicated edge backgrounds have been a common problem for algorithms that use an edge density approach, while the new localization method has a success rate of 86.4% based on a database of 822 images. While at this point in time the prototype system's LP recognition accuracy at 12.3% is not practical, several of its rudimentary technique employed could be exchanged for more sophisticated and effective ones and thus improving system performance significantly in future efforts. However, this work demonstrates that the de-skewing process can be significantly advantageous. TR: Li,R., Yassin-Fort, M.Y., and Anagnostopoulos G.C. (2008) Multi-stage Automatic License Plate Location & Recognition, Technical Report TR-2008-02, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Real-time, Static and Dynamic Hand Gesture Recognition for Human-Computer Interaction By: S.M. Hassan Ahmed , Todd C. Alexander Graduate mentor(s): Faculty mentor(s): Dr. Georgios C. Anagnostopoulos Abstract: Real-time, static and dynamic hand gesture recognition affords users the ability to interact with computers in more natural and intuitive ways. The hand can be used to communicate much more information by itself compared to computer mice, joysticks, etc. allowing a greater number of possibilities for computer interaction. The purpose of this work was to design, develop and study a practical framework for real-time gesture recognition that can be used in a variety of human-computer interaction applications with the aim of developing a prototype system for controlling Microsoft PowerPoint&tm; presentations. The approach taken is to locate hands starting with the finger tips after investigating potential regions of interest that are maintained through motion detection and subsequent tracking. The fingertips are very distinguishable from many backgrounds allowing a very robust tracking system. Using techniques such as Features from Accelerated Segment Test (FAST) corner detection the tips can be found efficiently. Circles generated using Bresenham’s algorithm were employed for finding corners as well as the center of the palm. Using features generated from these results along with a semi-supervised Adaptive Resonance Theory (ART) neural network, static gestures were able to be classified with an overall accuracy of 75%. Dynamic gestures on the other hand were able to be detected using the tra jectory formed by the center of the hand over a finite amount of time. Simple decision heuristics were then utilized to detect the movement of the hand. This research produced a working prototype of a robust hand tracking and gesture recognition system that can be used in numerous applications. TR: Ahmed, S.M.H, Alexander, T.C., and Anagnostopoulos G.C. (2008) Real-time, Static and Dynamic Hand Gesture Recognition for Human-Computer Interaction, Technical Report TR-2008-01, The AMALTHEA REU Program, Summer 2008 [PDF]
	Title: Testing and Improvement of the Triple Scoring Method for Applications of Wake-up Word Technology (2007) By: Andrew Stiles, Brandon Schmitt & Tad Gertz Graduate mentor(s): Tudor Klein Faculty mentor(s): Dr. Veton Kepuska Abstract: Constant monitoring of an individuals voice and near perfect recognition of a specific word while maintaining consistent rejections of all other words can be realized by implementation of Wake-Up Word (WUW) Speech Recognition (SR) technology. The algorithm shown here has the potential to add robustness to even in a speaker independent environment, and provides much better results for the application of single word recognition when compared to current industry or academic standards such as Microsoft SAPI and HTK respectively. By implementing a Triple Scoring Method (TSM) implemented with Hidden Markov Models (HMM) in the feature domain the WUW modeling results are found to be far superior in single word recognition, providing a 15166.15% increase in correct recognition with Callhome corpus over HTK and a 1303.78% increase over Microsoft SDK. TR: Stiles, A., Schmitt, B., Gertz, F., Klein, T., and Kepuska, V.Z. (2007) Testing and Improvement of the Triple Scoring Method for Applications of Wake-up Word Technology, Technical Report TR-2007-04, The AMALTHEA REU Program, Summer 2007 [PDF]
	Title: NEAT Drummer: Interactive Evolutionary Computation for Drum Pattern Generation (2007) By: Amy Hoover Graduate mentor(s): Faculty mentor(s): Dr. Kenneth Stanley Abstract: A major challenge in computer generated music is breaking the barrier between musical novelty and musical quality. Typically, computer music generators produce either genre-specific patterns that lack innovation or patterns that are given too much freedom and lack cohesion. In an attempt to both constrain the musical search space and produce novel rhythms, a program called NEAT Drummer is introduced. NEAT Drummer evolves neural networks with the NeuroEvolution of Augmenting Topologies (NEAT) that produce compelling drum patterns. To constrain the musical search space, NEAT Drummer accepts a base rhythm or motif from the user and through Interactive Evolutionary Computation (IEC), complexifies that pattern with each successive generation. This work discusses the concepts behind how NEAT Drummer understands and manipulates a base rhythm, which is either predefined by the user through a basic interface or defined by MIDI music file information. TR: Hoover, A.K., and Stanley, K.O. (2007) NEAT Drummer: Interactive Evolutionary Computation for Drum Pattern Generation, Technical Report TR-2007-03, The AMALTHEA REU Program, Summer 2007 [PDF]
	Title: A Grid Based System for Data Mining Using MapReduce (2007) By: Kelvin Cardona Graduate mentor(s): Jimmy Secretan Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: We discuss a Grid data mining system based on the MapReduce paradigm of computing. The MapReduce paradigm emphasizes system automation of fault tolerance and redundancy, while keeping the programming model for the user very simple. MapReduce is built closely on top of a distributed file system, that allows efficient distributed storage of large data sets, and allows computation to be scheduled closely to this data. Many machine learning algorithms can be easily integrated into this environment. We explore the potential of the MapReduce paradigm for general large scale data mining. We offer several modifications to the existing MapReduce scheduling system to bring it from a cluster environment to a campus grid that includes desktop PCs, servers and clusters. We provide an example implementation of a machine learning algorithm (the Probabilistic Neural Network) in MapReduce form. We also discuss a MapReduce simulator that can be used to develop further enhancements to the MapReduce system. We provide simulation results for two new proposed scheduling algorithms, designed to improve MapReduce processing on the grid. These scheduling algorithms provide increased storage efficiency and increased job processing speed, when used in a heterogeneous grid environment. This work will be used in the future to produce a fully functioning implementation of the MapReduce runtime system for a grid environment, that will enable easy, data intensive parallel computing for machine learning, with little to no additional hardware investment. TR: Cardona, K., Secretan, J., Georgiopoulos, M. and Anagnostopoulos G.C. (2007) A Grid Based System for Data Mining Using MapReduce, Technical Report TR-2007-02, The AMALTHEA REU Program, Summer 2007. [PDF]
	Title: A Backward Adjusting Strategy for the C4.5 Decision Tree Classifier (2007) By: Maria Garcia & Jason Beck Graduate mentor(s): Mingyu Zhong Faculty mentor(s): Prof. Michael Georgiopoulos Abstract: In machine learning, decision trees are employed extensively in solving classification problems. In order to produce a decision tree classifier two main steps need to be followed. The first step is to grow the tree using a set of data, referred to as the training set. The second step is to prune the tree; this step produces a smaller tree with better generalization (smaller error on unseen data). The goal of this project is to incorporate an additional adjustment phase interjected between the growing and pruning phases of a well known decision tree classifier, called the C4.5 decision tree. This additional step reduces the error rate (generalization of the tree) by making adjustments to the non-optimal splits created in the growing phase of the C4.5 classifier. As a byproduct of our work we are also discussing of how the decision tree produced by C4.5 is affected by the change of the C4.5 default parameters, such as CF (confidence factor) and MS (number of minimum split-off) cases, and emphasizing the fact that CF and MS parameter values, different than the default values, lead us to C4.5 trees of much smaller size and smaller error. TR: Beck, J.R., Garcia, M.E., Zhong, M. Georgiopoulos, M., and Anagnostopoulos G.C. (2007) A Backward Adjusting Strategy for the C4.5 Decision Tree Classifier, Technical Report TR-2007-01, The AMALTHEA REU Program, Summer 2007 [PDF]