ABSTRACT. This talk introduces a method for analyzing time series topic transition based on micro-clusters to present different situations that show people’s reactions to topical problems on the Web. To mine micro-clusters, we leverage our original algorithm, data polishing. Quality clusters are extracted as topics by the algorithm, so that the semantics of clusters can be easily analyzed. In the presentation, we show the dynamics of fake news using real tweets targeting the petrochemical complex explosion that happened just after the Great Japan East Earthquake in 2011.

ABSTRACT. This paper presents a new method for hand segmentation from images and video. The method based mainly on an advanced technique for instance segmentation (Mask R-CNN) which has been shown very efficient in segmentation task on COCO dataset. However, Mask R-CNN has some limitations. It works on still images, so cannot explore temporal information of the object of interest such as dynamic hand gestures. Second Mask R-CNN usually fails to detect object suffered from motion blur at low resolution as hand. Our proposed method improves Mask R-CNN by integrating a Mean Shift tracker that tracks hands in consecutive frames and removes false alarms. We have also trained another model of Mask R-CNN on cropped regions extended from hand centers to obtain a better accuracy of segmentation. We have evaluated both methods on a self-constructed multi-view dataset of hand gestures and show how robust these methods are to view point changes. Experimental results showed that our method achieved better performance than the original Mask R-CNN under different viewpoints

ABSTRACT. Thai License Plate Recognition is widely applied to identify car 's owner. In Thailand, police and soldier use TLPR to extract information on license plate for protecting car bomb in the South of Thailand. Currently,TLPR has low accuracy rate on province part. This research aims to apply Convolution Neural Networks to recognize province name on Thai license plate. We use dataset which are collected from real situation, each image has low resolution. This research, we recognized province name using Convolution Neural Networks, it obtains accuracy at 87.30%.

ABSTRACT. Predicting court judgement has gained growing attention over the past years. Prior attempts employed traditional prediction techniques based on Bag of words (BoW), where the order of words is discarded, resulting in low accuracy. In this paper, we propose a prediction model of criminal cases from Thai Supreme Court using End-to-End Deep Learning Neural Networks. Our model imitates a process of legal interpretation, whereby recurrent neural networks read the fact from an input case and compare them against relevant legal provisions with the attention mechanism. The model’s output shows if a person is guilty of a crime according to the fact and laws. After the performance test, we find that our model could yield the higher F1 than traditional text classification techniques including Naïve Bayes and SVM. In addition, we innovate the open dataset called “Thai Supreme Court Cases (TSCC)” that was compiled from many decades of Thai Supreme Court criminal judgments. It features the text of fact expertly extracted from each judgement, textual provisions from Thai Criminal Code, and binary-format labels following the theoretical criminal law structure. This dataset is useful for achieving judgement predicting task together with emulating actual criminal case trial.

ABSTRACT. Nowadays, the problems that arise within the city due to the increasing population. It is a major problem for cities around the world. A smart city is a way to make the city more efficient, sustainable and livable. To be a city that can be monitored and combine the functionality of all the important things. It can control and manage resources, infrastructure security, and maintenance. This research describes the meaning and concept of making a city to smart city and explore the facets and dimensions of a smart city to understand the concept of a smart city that is involved in which. Finally, talk about how to develop a smart city in Thailand and how will it work in the future.

ABSTRACT. Most of the snake mimicking robots achieve various gaits motion through periodic defined movements of all units, controlled by a single electronics processing unit which demands higher specifications and battery voltage. The motion achieved by sliding against surface results in improper imitation of snake motion due to limitation in speed of oscillations of motors and sliding on unpredictable surfaces. We propose a module and its prototype, which is able to achieve required motion on its own, with its batteries and processing unit placed inside it, and can operate independently. This module when multiplied and connected in different configurations can be utilized to develop robots like snake, hexapod, self-balancing robot, etc.

ABSTRACT. Unmanned Aerial Vehicle (UAV) has received a lot of consideration in many applications nowadays. Modern UAVs have been designed not only for performance but for efficiency as well. Propeller is an important part for UAV performance. However, the wake from the propeller or propeller induced-flow could interfere with the UAV aerodynamics. For UAV design and aerodynamics analysis, the computational fluid dynamics (CFD) method is a cost-efficient, popular method. But, the full flow simulation of UAV with actual propeller model can be quite complex, time-consuming, and require high computational resource. In this study, a virtual rotor model is deployed in CFD environment in order to computationally-efficiently simulate the airflow generated by UAV propeller. The virtual rotor considers the propeller disk as a surface of pressure difference. The parameters of the virtual rotor are obtained for the propeller approximation model developed in previous work [1]. In this paper, the CFD simulations of wake (airflow velocity) from a commercially-available propeller using its actual propeller (3D-scan) model and the virtual model in a hovering condition are investigated. The CFD simulations using actual propeller model and virtual model show an error of 15.5% and 10% from averaged thrust of static test data at 6000 RPM, respectively. The comparison results show an agreement between airflows of the actual and virtual models.

ABSTRACT. Development of Unmanned Aerial Vehicle (UAV) for defence application continues for several application in recent years. Quadrotor as, an example of UAV, have been designed not noly for manuvering but also for surveillance purpose, which include video recording and data transmitions. Event-triggered, an alternative strategy to allocate bandwidth in data communication, is introduced to meet data traffic requirement. This mechanism makes controlling only happen when certain condition is fullfilled. By doing so, bandwidth data can be used for other process. This paper presents a method to control of a quadrotor using event-triggered scheme, in which the LQR is applied. Comparison study between the scheme and without event-triggered are conducted to see how communication is used. Result shows that event-triggered do decrease number of communication. It is also observed that peformance of quadrotor under event-triggered no longer as good as classical control. because of that, this communication and peformance trade off should be considered when implementing event-triggered.

ABSTRACT. The basic performance of vehicles such as acceleration, deceleration, steering is determined by the interactive force generated between the tire and the ground, which is an important factor for running vehicles. A friction coefficient is one of the important indicators representing the interaction, and it determines the accelerating and braking performance. Also, the relationship between tire sideslip angle and the side force greatly affects the turning performance of the vehicle. These interactions are different depending on the ground such as roughness of the surface, dry or wet, etc. Especially for vehicles running off-paved roads, the interaction varies depending on the soil that the vehicle travels. It also changes greatly due to vegetation, roughness of the ground, and other factors. For these reasons, it is necessary to measure the condition of the road surface during traveling to clarify the characteristics of the tire and the ground. This interaction problem is also important, for controlling the vehicle in the case of unmanned operation. This report shows that valuable information of the interaction between a tire and the ground can be obtained by processing the data, which includes forces on tire, the tire rotational and travel velocities, and the posture angles of the tire, collected through the tire measurement system during the vehicle runs on actual grounds. The authors discuss the way of processing data and the feasibility of obtaining the interaction information using these data on actual grounds.

ABSTRACT. This talk consists of two parts. First to discuss the decision making in a digital economy and a digital society, as well as the relationship between AI and recent ICT breakthrough including data science. Second to illustrate the power of data science through problems and solutions in medicine, transportation, customer relationship, and more.

ABSTRACT. A tactical wireless network is a military radio communication network supporting mission-critical applications. Hence, a tactical wireless network demands more reliability, availability, robustness, and security than a commercial wireless network. The tactical wireless network must operate in hostile environment, where the environment changes rapidly and is prone to attack. To maintain the required quality of services (QoS), the network must intelligently adapt to the hostile environment. The concept of cognitive radio (CR), in which a radio can sense and adapt to radio environment, could be a solution for modern tactical wireless networks. Machine learning plays an important role in CR to provide sensing and adapting functions. Irrational decision made by a machine learning can lead to flaws in the CR. The introduction of Deep Learning models machine learning on the basis of human brain process; and hence, could make the CR more rational. This paper explores the challenges of tactical wireless networks, the CR functions as the solution for tactical wireless networks and Deep Learning techniques for improving CR functions. The survey presented in this paper should contribute to the development of modern tactical wireless networks by providing the possible applications, benefits and drawbacks of deep learning in CR.

ABSTRACT. A real-time decision support system with the ca- pability to provide information related to possible criminal’s escape path can be very useful for a law enforcement to pursue a perpetrator after a crime has been committed. Typically, the exact escape path is unknown, and pursuers must relied on a predicted path based on available information about the environment. In static environment, a perpetrator may escape through an optimal path that is predicted using any existing optimal path finding algorithms. However, the path can be dynamic when environment is changed. The perpetrator may decide to change path when there is information about the environment up ahead. This paper models the perpetrator’s path selection as a Markov Decision Process (MDP) and apply Q-learning to solve for a perpetrator’s escape path. The experiment results shows that our algorithm can find most probable escape path in the dynamic environment, which can be significant reference in a real-time decision support system for law enforcement applications.

ABSTRACT. Early detection of new kinds of malware always plays an important role in defending the network systems. Especially, if intelligent protection systems could themselves detect an existence of new malware types in their system, even with a very small number of malware samples, it must be a huge benefit for the organization as well as the social since it help preventing the spreading of that kind of malware. To deal with learning from few samples, term “one-shot learning” or “few-shot learning” was introduced, and mostly used in computer vision to recognize images, handwriting, etc. An approach introduced in this paper takes advantage of One-shot learning algorithms in solving the malware classification problem by using Memory Augmented Neural Network in combination with malware’s API calls sequence, which is a very valuable source of information for identifying malware behavior. In addition, it also use some advantages of the development in Natural Language Processing field such as word2vec, etc. to convert those API sequences to numeric vectors before feeding to the one-shot learning network. The results confirm very good accuracies compared to the other traditional methods.

ABSTRACT. The purpose of this paper is to investigate the ability of binary classification prediction using the mining of possitive negative association rules (PNARs) for input data tuples of continuos or binary categorical variables where the number of the variables is very large. In the paper, the PNARs (as understood until now) called narrow positive negative association rules (NPNARs for short) as well as their generalized forms called extended positive negative association rules (EPNARs for short) will be investigated. The paper proposes a classification algorithm based on mining the discriminant NPNARs and EPNARs in which the consequent part consists of only a item or this item with negation. The algorithm integrates Dimensionality Reduction, Rule Generation, Removal of redundant rules and Classification. The number of classified instances in testing datasets and the classification performance using the valid disriminant NPNARs and EPNARs found out by the proposed algorithm are compared with each other and with 12 other classification algorithms based on decision trees, on Bayes classifier, on functions, on rules and on "meta" algorithms via the experiment of these algorithms on the actual datasets in the UCI database. Experimental results show that the number of classified instances using the mining of NPNARs is the lowest whereas using the mining of EPNARs is generally higher than that by the reported algorithms, but not always the case. The classification accuracy using the mining of NPNARs is always approximated by the highest classification accuracy performed by one of the reported algorithms. The accuracy of classification using the mining of EPNARs is quite high and in general it is lower than the highest accuracy but this is not always true. The paper also shows that the mining of EPNARs rather than that of NPNARs has prospects to become a classification technique applied in the real world.The paper also outlines some issues that need to be researched in the near future so that using the mining of the discriminant EPNARs becomes an efficient method for classifying high dimensional datasets.

ABSTRACT. This research study relationships between fraud and non-fraud transactions of the ATM. Given the transactions, which are obtained from a commercial bank in Thailand, a behavior of each account in one day such as the total amount withdraw and the number of transactions are extracted. Given the set of transactions in each account, the reference is computed within the specific time-window. The reference value and spread are used for constructing a boundary to define the normal and abnormal behavior. Given an instance of aggregate transactions over certain time window, it is predicted as a fraud if the total amount and the total number of transactions per day lies outside the boundary. Experiments are conducted to compare the accuracy between different aggregation time windows and boundaries with different spreading factors.

ABSTRACT. A distributed denial-of-service (DDoS) attack characterized by flooding SYN packets is one of the network attacks to make the information system unavailable. This kind of attack becomes dangerous and more difficult to prevent and defense when attackers try to send flood SYN packets with spoof source, especially, there packets have information fields as the normal SYN packets. In the previous study, we proposed method of PIDAD - Packet Identification Anomaly Detection to detect above mentioned spoofing packets. However, this method has some limitations which are in need of further research and completion. In this study, we suggest the method of using many layers of Bloom Filter to address the limitations of previous method we proposed before.

ABSTRACT. Early Effort Estimation in software development is an essential process in project management, especially in project planning, budget supervising and resource control. Accurate estimation in the early stages of software development can help planning the project more effectively and significantly reduce cost along with many other benefits, but these estimations can be arduous to calculate accurately because the required input and parameters which software estimations are based on are vague and the limit of time and resource. However, there are some techniques and models that can minimize or even exterminate imprecise parameters. In this paper, Function Points Analysis (FPA) technique and Adaptive Neuron Fuzzy Inference System (ANFIS) model have been chosen for this article. Dataset used for analysis is the Albrecht dataset (15x240) which has been divided into 3 types of Train-Test Dataset, separated by groups of 5 or 6 General System Characteristics with a range of 70-120 rows. These Train datasets will be trained by ANFIS to calculate result in Test datasets. The ANFIS is modeled for 6 membership functions which will be experimented in this article are Gaussian curve, Generalized bell-shape, Triangular-shaped, Gaussian combination, Pi-shaped and Trapezoidal-shaped membership function. The performance of the proposed models will be evaluated with other model parameters to evaluate several criteria such as Mean Relative Error (MRE), Mean Magnitude of Relative Error (MMRE). According to the results, the Gaussian curve membership function has the lowest MMRE in all the proposed membership functions. The result is also the improvement with higher accuracy, in compare with Function Point Counting Practices Manual 4.2.1 formula.

ABSTRACT. The global reach of IoT and the diversity of IoT cultures and practices introduces multiple confusions that blur the understanding of the Internet of Things domain. This confusion is now amplified by the technical-advertising bubble created around the latest generation of "connected objects". This motivates the creation of a reference model for the field of the Internet of Things in order to promote a common understanding of its concepts, objects and their relationships. This reference model will then be adopted to constitute an IoT domain specific modeling language. This language will be, then, used to propose a development approach around it. This article attempts to present the reference model of IoT, the IoT domain specific modeling language based on it and the development approach around it. The approach through the IoT domain specific modeling language allow a developer to model his IoT solution. Next, he can generate a Java code from this solution Model.

ABSTRACT. Military training is crucial for the success of a military mission. The Royal Thai Army Aviation Center requires an attack helicopter (AH) simulator to address the limited opportunity for live training and the shortage of skilled AH operators. The initial development effort focuses on a physical mock-up of the attack helicopter and training in the operation of the gunner instrument panel. This paper details our new development to improve the simulator in both hardware and software aspects. We apply GIS and remote sensing techniques to create a large-scale geo-database of a real training area. We import the geo-database into a game engine to create a realistic virtual terrain and an interactive training scenario that can be used to teach relevant military strategies and tactics, and to provide context for the instrument panel training. We demonstrate the simulator to military experts. The feedback suggests the simulator can be a useful tool for training and can increase trainee confidence. We consider suggestions for improvement and identify additional ideas for future investigations.

ABSTRACT. Let $G = (V, E)$ be an undirected graph, where $V$ is the vertex set and $E$ is the edge set. A subset $M$ of $E$ is an \emph{induced matching} of $G$ if $M$ is a matching of $G$ and no two edges in $M$ are joined by an edge. Finding a \emph{maximum induced matching} is a $\mathbb{NP}$-Hard problem on general graphs, even on bipartite graphs. However, this problem can be solved in polynomial time in some special graph classes such as weakly chordal, chordal, interval and circular-arc graphs. In this paper, we introduce a maximum induced matching algorithm in permutation graphs with $\mathcal{O}(|V| k(G) \log \log(|V|))$ time in worst case complexity and $\mathcal{O}(|V| \sqrt{|V|} \log \log(|V|))$ time in average case complexity, where $k(G)$ is the cardinality of the minimum clique cover set. The approach is to reduce the size of vertex set of $L(G)^2$ without changing the cardinality of its maximum independent set. Our algorithm has much better time complexity than the best known algorithm in both worst case and average case.