Real-time Optimal Selection of Multirobot Coalition Formation Algorithms using Conceptual Clustering

Intelligent multirobot coalition formation systems equipped with mission driven algorithm selection strategies are crucial for a wide spectrum of real-world situations. i-CiFHaR is an unique coalition formation framework that comprises a library of diverse algorithms and uses a taxonomy-based real-time probabilistic reasoning to select the most appropriate algorithm(s) to apply. However, i-CiFHaR suffers from high computational time with an increase in the number of algorithms. This paper is the first to demonstrate the use of conceptual clustering in order to mine crucial patterns and relationships among existing coalition formation algorithms. iCiFHaR leverages the derived optimal hierarchical classification tree to analyze only the most appropriate algorithms’ cluster for application to a real-world mission scenario. The results show that the conceptual clustering technique reduces computation time by 67%. The multirobot coalition formation problem seeks to intelligently partition a team of heterogeneous robots into coalitions for a set of real-world tasks. Besides being NPcomplete (Sandholm et al. 1999), the problem is also hard to approximate (Service and Adams 2011a). Traditional approaches to solving the problem include a number of greedy algorithms (Shehory and Kraus 1998; Vig and Adams 2006b), approximation algorithms (Sandholm et al. 1999), and auction-based approaches (Vig and Adams 2006a; Gerkey and Matarić 2002; Shiroma and Campos 2009). A single greedy algorithm-based system generates coalitions quickly, but fails to guarantee solution quality. Despite guaranteeing solution quality, approximation algorithm-based systems are inappropriate for real-time applications with large teams of robots, because of their high worst case run-time complexities. Auction-based algorithms provide the desired scalability and decentralization, but are inadequate for low communication environments due to their high communication overhead. Therefore, conventional single algorithm-based systems are inadequate and brittle for a wide spectrum of complex, uncertain real-world missions. These limitations led to the development of the intelligent-Coalition Formation framework for Humans and Robots (i-CiFHaR) that incorporates a library of algorithms, Copyright c © 2015, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. each applicable to different categories of real-world problems (Sen and Adams 2014). The framework exhibits probabilistic reasoning in order to make online decisions regarding the most suitable algorithm(s) to apply based on multiple mission criteria. The intelligent, real-time selection of appropriate algorithms from the library via a decision network renders i-CiFHaR flexible and adaptive, when computing robust coalitions for a wide spectrum of real-world missions. Despite the successful demonstration of i-CiFHaR’s algorithm selection capability for a wide variety of mission scenarios, it suffers from a high computational time as the number of algorithms and taxonomy attributes increase. The presented work mines patterns in i-CiFHaR’s algorithms using conceptual clustering, an unsupervised machine learning approach in order to extract the most suitable cluster of algorithms for application analysis in a specific mission; thereby, accomplishing better scalability and reduced computational time. The presented framework is the first to leverage a conceptual clustering technique to partition any set of coalition formation algorithms in order to derive an optimal hierarchy classification tree, given any classification taxonomy. The results contribute to the state-ofthe-art in multiagent systems by demonstrating the existence of crucial patterns and intricate relationships among existing coalition algorithms. The presented i-CiFHaR framework leverages COBWEB, a conceptual clustering algorithm (Fisher 1987) for identifying clusters of similar algorithms in the library. Rather than employing all the algorithms, many of which may not be applicable during a particular mission situation, the improved framework uses the probabilistic metric, category utility (Gluck 1985) to identify the most suitable cluster of algorithms. Based on the selected cluster, the decision network optimizes the ranking of the algorithms in the chosen cluster by maximizing the expected utility scores. The experimental results show that algorithm rankings match those found for each of the twenty four mission scenarios as the original i-CiFHaR, but requiring approximately 67% less computation time. The Background section provides a comprehensive overview of the related work. The incorporation of the conceptual clustering technique is described in the System Design. The experimental design, results, and the conclusions are provided in the subsequent sections.