Novel approaches to cooperative coevolution of heterogeneous multiagent systems

Tese de doutoramento, Informática (Engenharia Informática), Universidade de Lisboa, Faculdade de Ciências, 2017Heterogeneous multirobot systems are characterised by the morphological and/or behavioural heterogeneity of their constituent robots. These systems have a number of advantages over the more common homogeneous multirobot systems: they can leverage specialisation for increased efficiency, and they can solve tasks that are beyond the reach of any single type of robot, by combining the capabilities of different robots. Manually designing control for heterogeneous systems is a challenging endeavour, since the desired system behaviour has to be decomposed into behavioural rules for the individual robots, in such a way that the team as a whole cooperates and takes advantage of specialisation. Evolutionary robotics is a promising alternative that can be used to automate the synthesis of controllers for multirobot systems, but so far, research in the field has been mostly focused on homogeneous systems, such as swarm robotics systems. Cooperative coevolutionary algorithms (CCEAs) are a type of evolutionary algorithm that facilitate the evolution of control for heterogeneous systems, by working over a decomposition of the problem. In a typical CCEA application, each agent evolves in a separate population, with the evaluation of each agent depending on the cooperation with agents from the other coevolving populations. A CCEA is thus capable of projecting the large search space into multiple smaller, and more manageable, search spaces. Unfortunately, the use of cooperative coevolutionary algorithms is associated with a number of challenges. Previous works have shown that CCEAs are not necessarily attracted to the global optimum, but often converge to mediocre stable states; they can be inefficient when applied to large teams; and they have not yet been demonstrated in real robotic systems, nor in morphologically heterogeneous multirobot systems. In this thesis, we propose novel methods for overcoming the fundamental challenges in cooperative coevolutionary algorithms mentioned above, and study them in multirobot domains: we propose novelty-driven cooperative coevolution, in which premature convergence is avoided by encouraging behavioural novelty; and we propose Hyb-CCEA, an extension of CCEAs that places the team heterogeneity under evolutionary control, significantly improving its scalability with respect to the team size. These two approaches have in common that they take into account the exploration of the behaviour space by the evolutionary process. Besides relying on the fitness function for the evaluation of the candidate solutions, the evolutionary process analyses the behaviour of the evolving agents to improve the effectiveness of the evolutionary search. The ultimate goal of our research is to achieve general methods that can effectively synthesise controllers for heterogeneous multirobot systems, and therefore help to realise the full potential of this type of systems. To this end, we demonstrate the proposed approaches in a variety of multirobot domains used in previous works, and we study the application of CCEAs to new robotics domains, including a morphological heterogeneous system and a real robotic system.Fundação para a Ciência e a Tecnologia (FCT, PEst-OE/EEI/LA0008/2011