Adaptive Multi-agent System for Dynamic Clustering Applied to Itineraries Regularities and Traffic Prediction

Alexandre Perles, Ha-Nhi Ngo, Elsy Kaddoum, Valérie Camps

Abstract: Nowadays, electronic devices such as mobile phones, sensors embedded in vehicles, and more generally digital acquisition devices continuously provide information relating the state of the surrounding environment at nearly real-time. These data provide important information required for real-time systems monitoring such as temperature regulation in smart building, traffic planning to relieve network congestion in smart cities, etc. The growing demand for analysing such data encourages researchers to adopt an approach known as “streaming analysis” which aims at processing data streams at real-time to continuously capture data evolution over time. Data stream clustering approaches already exist but require keeping in memory all the input data and have a slow adaptation to data changes. To solve these problems, we propose to agentify the clusters and allow them to fuse and evolve locally and autonomously. This work presents AMAS4DC a generic Dynamic Clustering model based on Adaptive Multi-Agent System approach. AMAS4DC processes acquired data on the fly using local similarity evaluation for cluster’s creation or fusion. AMAS4DC is then instantiated on two use cases: the dynamic clustering of itineraries to detect regularities and the dynamic clustering of traffic data to predict future traffic. The conducted experiments underline the performance of AMAS4DC in terms of memory usage, processing time and clustering quality compared to well-known models for dynamic clustering.