Building an Multi-Agent Whisky Recommender System

Abstract

MAS (Multi-Agent Systems), classifiers and other AI (Artificial Intelligence) techniques are increasingly becoming more common. The capability to handle complex and advanced problems by MAS was explored in this thesis. An MAS duty-free shopping recommender system was built for this purpose. The MAS system was part of a larger system built by the AmbieSense project. In addition, the AmbieSense project had built a prototype that was tested at Oslo Airport (OSL) Gardermoen. As a test case, the duty-free shopping system was set to classify and recommend whiskies. The system incorporated several AI techniques such as agents, ontology, knowledge base and classifiers. The MAS was built using the JADE-LEAP framework, and Protégé was used for constructing the knowledge base. Various tests were performed for testing the system. Firstly, the agent communication was monitored using a Sniffer Agent. Secondly, the system’s ability to run on mobile devices was tested on a PDA and various mobile phones. Thirdly, the MAS abilities compared to a ‘normal’ computer program were tested by replacing agents at run-time, using several JADE platforms, and by the experience gathered during development and the use of the developed system. Lastly, the recommendation was cross-validated against Dr. Wishart’s whisky classification. Weka was employed as a tool for testing different features and classifiers. Different classification algorithms are explained such as NNR (Nearest-Neighbour Rule), Bayesian, CBR (Case-Based Reasoning), cluster analysis and self-organizing feature maps. Two classification algorithms were tested; NNR and Bayesian. Features were tested using feature evaluation algorithms; information gain and ReliefF. The accuracy of the classification was tested using 10 fold cross-validation. The testing showed that it is possible to make an MAS handling complex and advanced problems. It has also been shown that an MAS have some benefits in the areas of reliability, extensibility, computational efficiency and maintainability when compared to a ‘normal’ program. The error rate produced by the classifier was 56%; a figure which is too high for a recommendation system. Improvements could probably be achieved by finding better features or by selecting a different classifier. The system developed does not necessarily have to be used for duty-free shopping but could also be used for any shopping items.