Proceedings of the 10th Annual Sessions, Sri Lanka Association for Artificial Intelligence

29th November 2013


A Fuzzy-Mathematical Model to Recover Motion with Monocular Vision

S. Pathirana, J. G. Samarawickrama, A. S. Karunananda

Abstract-This paper describes a fuzzy based mathematical model to recover the motion path of an obstacle observed via a single camera. The strategy is analysing a sequence of images captured in regular time intervals, more specifically, studying the variation of the apparent size of an obstacle and the relative position change on reference frame. Those two measures are the only inputs to the fuzzy-mathematical model, the major emphasis of the research, which recovers the motion path as an equation. Necessary feature extraction is being achieved via a real time image processing module which relies on an optical flow technique as the key technique to recognize dynamic objects. It was reported a 91.98% average accuracy from the fuzzy-mathematical model in simulation environments, where the inputs were generated by a simulator program in order to study the precision of the fuzzy-mathematical model as a standalone application. An average accuracy of 59.8% was experienced at the real time application, an artefact to test the postulated concept in real time dynamic environment, which comprised of three major modules: a real-time image processing module, the fuzzy mathematical model and a mobile robot.



MaSIO – a Solution to Information Overflow in Agriculture

B. Hettige, A. S. Karunananda, G. Rzevski
Abstract-Information overflow has been a hindrance to search for information even by computing experts and subject matter experts. The situation is worst with the persons who have limited computing knowledge and unfamiliarity in the relevant subject. We argue that technological advancements in computing could provide solutions for the said issue.  Our research has been inspired by the power of Multi Agent Systems (MAS) technology to model complex systems encompassing large number of distributed and interconnected entities that might change over the time. We have developed a Multi Agent System, MaSIO, to address the above issue. MaSIO mimics a scenario where an IT expert communicates with reliable subject matter experts and find information. In MaSIO, the roles of an IT expert and the subject matter experts have been implemented as Agents. MaSIO has been implemented as a Java-based generic solution. This paper presents customization of MaSIO for the domain of Agriculture. The paper also demonstrates how MaSIO extract reliable information regardless of the level of computing literacy of user.



Modeling of Hidden Layer Architecture in Feed-forward Artificial Neural Networks

N. M. Wagarachchi, A. S. Karunananda
Abstract-Determining the appropriate architecture of a neural network is one of the main unsolved problems in artificial neural networks. The architecture   has a great impact on its generalization power.  More precisely, by changing the number of layers and neurons in each hidden layer, generalization ability can be significantly changed.  Therefore, the architecture is crucial in artificial neural network and determining the hidden layer architecture has become a research challenge.  In this paper, a pruning technique is presented to obtain an appropriate architecture by using the delta values of hidden layers.  Pruning is done by using the delta values of hidden layers.  The proposed method has been tested with three benchmark problem datasets in artificial neural networks and machine learning namely, breast cancer, Iris and car evaluation.  The experimental results show that the modified architecture with lesser number of neurons performs better in generalization than that of the back-propagation algorithm.



WMAC: Web-Based Multi-Agent Solution for Agriculture Community

H. M. H. R. Jayarathna, B. Hettige

Abstract-Multi-agent System Technology (MAS) is one of the powerful technologies which is used to solve real world problems. Compared with existing Multi-agent applications, web based solutions are more useful than the standalone applications. This paper presents web based multi agent system named WMAC which can be used to communicate with the people engaged in Agriculture industry. Web based multi-agent system uses a common MYSQL database as the ontology of  each agent.  Agents work as web clients and design through the PHP and AJAX technologies. System provides four types of agents as farmer, buyer, seller and instructor; which represent farmers, buyers, sellers and technical instructors in the agricultural community and which makes the communication among persons in the agricultural industry as required. The WMAC is a web based development of existing Java based standalone multi-agent system named AgriCom. The WMAC system has been successfully tested with the practical environment and successful results were achieved.



Multi Agent System for Artificial Neural Network Training

S. M. Dharmakeerthi, A. S. Karunananda

Abstract-Artificial neural networks are heavily used in the areas of pattern recognition, feature extraction, function approximation, scientific classification, control systems, noise reduction and prediction. Feed-forward and back-propagation neural networks are the most commonly used artificial neural networks. Many researchers face difficulties when selecting a proper ANN architecture and training parameters. The manual ANN training process is not the best practical solution because it is a time consuming task. Also most of the people conduct the manual process in an ad-hoc manner without having a proper basis for changing parameters. This research project has developed a multi-agent based approach to automate the optimization of neural network architecture and its training for feed-forward and back-propagation neural network. The ontology of the agent system comprises of commonly used heuristics for training of neural networks. Our experiments show that the more rational results can be obtained from the system with both simple datasets like XOR as well as with real life data sets. We can conclude that the neural network optimization and training tasks can be successfully accomplished by the agent based approach by analysing the results of the evaluation.



Conflict Reduction Analysis of Bulk Agent Approach in Multi Agent Systems

P. M. Gunathilaka, A. S. Karunananda

Abstract-Our universe can be considered as the largest multi agent system with no visible conflicts. Particles in different dimensions interacts, based on different gravitational rules, which defines the universal extra dimensions called ‘Bulk’. The same concept can be modeled, as the Bulk Agent Approach in multi agent systems to overcome potential conflicts, which also empowers the direction of the emergent success of the overall system. On the other hand, it is a design challenge in multi agent systems, how to avoid unnecessary conflicting chaos, which could consume large computational resources and valuable time. Lack of resources or social knowledge could lead to either resource conflicts or knowledge conflicts. As a solution, Argumentation Based Negotiation (ABN) with the support of conflict evading and re-planning has been presented in the literature as one of the best approach in conflict resolution techniques. However, conflict evading and re-planning would not be useful in an environment where resources are not abundant. Therefore, we present our novel approach as a solution for the burning limitations of conflict evading and re-planning. Philosophical explanations and Brane Cosmology, which explains on how gravity governs on brane particles, based on the concept of universal extra dimensions, are the main inspirations for our research. Any multi agent environment can be considered as a multi-dimensional universe, where the universal norms originate in a higher dimension. These universal norms provide the guidelines for emergent success of the whole system. However, universal norms can change dynamically based on the social and environmental changes in the lower dimensions. Therefore, in our architecture we define higher dimensions by an agent type called Bulk Agents whereas agents in the lower dimensions are called Brane Agents. The Bulk Agent monitors behaviors of the Brane Agents and provides the direction or the guideline for the success of the overall system. These directions were shared among brane agents as Volatile Ontology so that the overall agent society is well capable of avoiding potential conflicts which otherwise would increase the failure rate of the system. Our analysis are experimented based on an application called Multi Agent Marketplace. Our experiments were analyzed based on statistical figures which has shown that the conflicts can be avoided or resolved with minimal computational time and resources by introducing Bulk Agents, which represents the extra dimensions in multi agent systems. This paper presents the results of our analysis on identifying the level of effectiveness of the Bulk Agent approach in conflict resolution in Multi Agent Systems. Keywords-multi agent systems; bulk agent; conflict resolution; brane agent; cosmology


Bone Crack Detector Based on X-Ray Using Fuzzy Logic and Neural Network

K. A.W.P. Abesinghe, S. D. H. S. Wickramarathna

Abstract-Large number of X-Ray images are analyzed by doctors in hospitals daily to identify various diseases   and anomalies in human body. One particular area is identification of problems in bones in that sense, based on X-Ray images doctors are going to predict the problem in that bone  such as bone crack, damage etc. These kinds of manual inspection of X-Ray images consume a lot of time and the process itself is monotonous made during the inspection.  As a solution to these problems here we introduce Computer-assisted decision-making system to detect cracks in bones which are visible in X-Ray image using Fuzzy Logic and Neural Networks.