A Blackboard based Hybrid Multi-Agent Approach for Machine Learning using Reinforcement Learning Techniques

  • Vasileios Manousakis Kokorakis

    Student thesis: Doctoral Thesis


    This thesis argues that diversity in error rate and heterogeneity in classifiers can successfully improve prediction performance through voting and then reward them for their success or failure. To that end, a central manager called the blackboard, it is capable of learning the area of expertise for each agent and exploit their capabilities to boost the predictive performance. We apply and combine several traditional ML methods and algorithms in a novel way in order to construct a blackboard based multi-agent system. In experimental work we focus on two case studies. The first case study includes datasets where accuracy is the performance metric that needs to be improved. On the other hand, the case study focuses on datasets for which recall, specificity and precision are the metrics boosted.

    Firstly, a novel blackboard based approach structure are introduced. An intelligent central manager exploits the knowledge diversity. The implementation and combination of the individual components described below, offer the ability to capture classifiers properties and quantify them using artificial intelligent techniques. Secondly, reinforcement learning techniques were utilized to reward its agent members with a trust value that reflects their competence. Thirdly, a number of computational voting functions are implemented and used in a novel way in order to let the agents express their opinion. Trust reward has an influence in each vote. Fourthly, two optimization methods are applied in order to optimize individual parameters included in the proposed trust function. Finally, three versions of the proposed approach are described and evaluated: naive, which is unaware of any information about the examined problem, then a version that is aware of certain patterns that are extracted during cluster analysis, and finally, a dynamic version that is capable of adapting dynamically.

    To sum up, this work has mainly contributed to knowledge by developing an intelligent multi-agent framework to exploit collective experience and improve the global prediction performance. To achieve that, three different version of the proposed approach, and a reinforcement learning based manager were introduced. At the same time, five voting schemes were implemented influenced from ta trust metric. The proposed approach is evaluated on a number of datasets from different domains e.g. bio-medical, Stack Exchange, SMS Spam and Financial.
    Date of AwardMar 2018
    Original languageEnglish
    Awarding Institution
    • University of Brighton
    SupervisorMiltos Miltos Petridis (Supervisor) & Stylianos Kapetanakis (Supervisor)

    Cite this