Proceedings of the 9th Annual Sessions, Sri Lanka Association for Artificial Intelligence
18th December 2012, Colombo
Multi-agent System Technology for Morphological Analysis
B. Hettige, A. S. Karunananda, G. Rzevski
Abstract-Machine Translation involves multiple phases including morphological, syntax and semantic analysis of source and target languages. Despite there are numerous approaches to machine translations, handling of semantics has been an unsolved research challenge. We have been researching to exploit power of multi-agent Systems technology for machine translation by extending our rule-based machine translation system, BEES. Since there are no agent development framework specific to machine translation, our project has started by developing our own framework, MaSMT. This paper presents our research on the development of morphological analysis phase in MaSMT. Twenty-two ordinary agents and one manager agent have been implemented to model morphological analysis of English language. In contrast, MaSMT implements 206 agents and a manager agent to handle morphological analysis in Sinhala language. MaSMT has been developed in Java, while BEES is a Prolog implementation. Performance of morphological analysis by MaSMT and BEES has been evaluated. It was revealed that MaSMT performs much faster than BEES for morphological analysis of English sentences with a reasonable length such as 15 words. In case of Sinhala language too, MaSMT performs better than BEES. The difference in performances of MaSMT in Sinhala and English reflects the number of morphological rules in two languages. Due to parallel execution, MaSMT shows a significant improvement in identification of syntactic categories of words that have more than one interpretation. This feature will be reflected even better in syntactic and semantic analysis as they necessarily involve rules with multiple interpretations.
Multi-Agents to Work with Primary Child:Expert Agents Architecture and Implementation
M.S. Nanayakkara, K.A. Dilini T. Kulawansa
Abstract-As an important part of today’s socio-techno-ecosystem, it is very vital to acknowledge the importance of primary education. Problems like fair distribution of resources, resources utilization, and microscopic concern on aspects related to the psychology of a child cause any primary education system, globally, face significant challenges when being implemented.CHILD@EDU is an alias given to a piece of research and development work which has being initiated to find solutions for these matters in primary education. CHILD@EDU fundamentally targets the children that involve in Primary Education whileproposing a knowledge system that addresses the vital and sensitive attributes of a common primary education system (known as the fundamental characteristics of a child with effective learning abilities). This paper is intended to publish the technical research, architecture, and implementation work related to the Expert Agents module of CHILD@EDU. This distributed multi-agent architecture has been tested in the means of project CHILD@EDUto be functioning with a higher accuracy adhering to the ontological expert decision support based on international WISC®-IV Assessment Standards.
A Multi-agent based Approach to Simulate Uncertainty of a Crowd in Panic
with Sharable Ontologies
Wimal Perera, A.S. Karunananda, Prasad Wimalaratne
Abstract-Crowd simulation is listed under many practical applications in computer industry; such as safety modeling, pre-planning building architectures, urban modeling and entertainment software. Most of these existing simulations are created by implementing computer algorithms based on extending deterministic models such as particle systems, clustering, cellular automata and fluid motion. However, extending a crowd model to simulate uncertainty of crowd behaviour during panic still remains a key challenge; since a computer algorithm approaches a solution by parameterizing predictable elements within a problem. It is evident from literature about the proven success of multi-agent technology behind modeling complex systems; comprising of many distributed entities interacting with each other and operated under lot of uncertainty. Thus it can be postulated that multi-agent technology provides a basis to model the uncertainty raised within a crowd during panic. Our proposed solution simulates this uncertainty by considering evacuation of a crowd from a building during fire. Each individual in the crowd is modeled as an agent associated with a local ontology. The local ontology of an agent is a collection of rules, representing the knowledge known to each individual prior to occurring fire. Rules embedded within local ontologies are shared among individuals as they interact with each other. As a result non-anticipated global behaviours arise within the crowd leading to emerging uncertainty during fire. Output of the system is a visualization of crowd behaviour during fire with recorded statistics. The statistics recorded during each simulation session indicate evacuation related information per each individual; providing a basis for evaluation by comparison with real world observations.
Agro Finance System
Manoj A.Bandara, Malith L. Karunaratne, D.D.M. Ranasinghe
Abstract-– Sri Lanka is mainly considered as anagriculture based country, yet the contribution from agriculture to the Gross Domestic Production (GDP)is very low at present. Nevertheless a larger portion of the Sri Lankan labor force is absorbed into agriculture sector. This clearly highlights that there is an issue in productivity in the agriculture sector and this has to be addressed immediately. One major factor that contributes to low productivity is due to lack of knowledge in identifying the most suitable crops to grow in a particular area. This project adopts a Multi – Agent based approach to get information on cultivation and productivity and Neural Network based approach to improve the financial level of farmers. The system is capable of predicting the cost of production as well as the expected price for a selected crop from the productivity enhancement module. The low values of root means square for the difference between the actual price and the predicted price indicate the high accuracy of the system.
Agent Negotiations for Improving Quality of Solutions from Multiple Perspectives
M.D.W Srimal, A. S. Karunananda
Abstract-Most of the real world problems can be solved using more than one method which may return slightly different solutions. For instance, statistical techniques, artificial neural networks, fuzzy logic and genetic algorithm can model the same real world problem subject to own strengths and weaknesses. However, it is evident that human beings can modify/improve solutions generated in the individual capacity through negotiations among the individuals. This concept has been employed in the Multi Agent Systems (MAS) technology which can model complex real world problems to achieve quality solutions beyond the individual capacity. In this work, MAS has been used to ensemble weather forecasting results individually generated by Artificial Neural Network (ANN) and Genetic Algorithms (GA) through negotiation among solutions. It considers ANN and GA as two agents. It has selected this application domain to demonstrate the concept since weather forecasting is important for many sectors such as agriculture, fisheries and transportation. Our MAS solution forecasts the rainfall for next twenty four hours with the use of set of present weather conditions as inputs for ANN and GA agents. The defined two agents are used to operate on an ANN and GA solutions that start negotiation & deliberation to produce a more rational forecasting. The experiment concludes that even when solutions by ANN Agent and GA Agent shows a disparity at the beginning, they reach to commonly agreeable solution through the negotiation in the multi agent solution with a 65% of success.
Framework for Discovery of Data Models using Genetic Programming
W.J.L.N Wijayaweera, A.S. Karunananda
Abstract-The field of Genetic Programming in Artificial Intelligence strives to get computers to solve a problem without explicitly coding a solution by a programmer. Genetic Programming is a relatively new technology, which comes under automatic programming. After the initial work by John R. Koza in genetic programming, much research work have been done to discover data models in various datasets. These work have been rather domain specific and little attention has been given to develop generic framework for modeling and experimenting with genetic programming solutions for real world problems. This paper discusses a project to develop a visual environment, named as GPVLab, to design and experiment with genetic programming solutions for real world problems. GPVLab has successfully discovered data models for various data sets and according to the main evaluation it is evident that GPVLab can generate solutions which provide better results in 56% of the time. It is concluded that GPVLab can be used to model genetic programming application very conveniently. GPVLab can be used not only for discovering data models but also doing various experiments in genetic programming.
Modeling Memory as Conditional Phenomena for a New Theory of Computing
W. A. C. Weerakoon, A. S. Karunananda, N. G. J. Dias
Abstract-Computers with the Von-Neumann architecture improve their processing power with the support of memory. This architecture has been improved by introducing different types of memories. The research we conduct has been inspired by the fact that humans are able to improve their memories with the support of continuous processing in the mind. It is evident that we can start with a smaller memory to drive the processing, which in turn improves, both the memory and the processing. This is analogous to how a person uses short-notes to do processing on a larger knowledge base. We postulate the processor to use the said smaller memory to access the bigger knowledge base, without the processor directly accessing the knowledge base as in the present computations. Thus we are researching into the development of the said small memory, say tactics memory, as a novel memory for Von-Neumann architecture. In doing so, we have exploited the Buddhist theory of mind, which presents everything as a phenomenon that occurs when the related conditions are met. We have developed algebra for modeling the tactics memory. The tactics memory can be introduced both as a software or hardware solution for computations.
Monocular Vision Based Agents for Navigation in Stochastic Environments
P.A.P.R Athukorala, A. S. Karunananda
Abstract-Autonomous navigation in a stochastic environment using monocular vision algorithms is a challenging task. This requires generation of depth information related to various obstacles in a changing environment. Since these algorithms depend on specific environment constraints, it is required to employee several such algorithms and select the best algorithm according to the present environment. As such modeling of monocular vision based algorithms for navigation in stochastic environments into low-end smart computing devices turns out to be a research challenge. This paper discusses a novel approach to integrate several monocular vision algorithms and to select the best algorithm among them according to the current environment conditions based on environment sensitive Software Agents. The system is implemented on an Android based mobile phone and given a sample scenario, it was able to gain a 66.6% improvement of detecting obstacles than using a single monocular vision algorithm. The CPU load was reduced by 10% when the depth perception algorithms were implemented as environment sensitive agents, in contrast to running them as separate algorithms in different threads.
Price Prediction Neural Network for Econ Centre
Prasad Wanigasinghe, D D M Ranasinghe
Abstract-Econ centre price prediction system works as a business intelligence system that practically helps wholesalers to predict the high marginal profit selling price of an item. According to the survey done based on “Dambulla” economic centre, increase or decrease of one rupee of an item can cause one wholesaler to gain or loss around 5000 rupees per day. Minimizing the loss by selecting the high marginal profit selling price is the main aim of this system. This aim was achieved by combining neural network time series analysis techniques and financial analysis techniques in cost accounting. This system is able to predict the required variables down to 1.01% error percentage. Using those predicted values the system then calculates the high marginal profit selling price and alternative prices.