Dynamics of Neuronal Models in Online Neuroevolution of Robotic Controllers

<strong>Dynamics of Neuronal Models in Online Neuroevolution of Robotic Controllers</strong>, by Fernando Silva, Luís Correia, and Anders Lyhne Christensen. To appear in: EPIA 2013. 16th Portuguese Conference on Artificial Intelligence. Springer-Verlag. <strong>Abstract: </strong>In this paper, we investigate the dynamics of different neuronal models on online neuroevolution of robotic controllers in multirobot systems.<br>We compare the performance and robustness of neural network-based controllers using summing neurons, multiplicative neurons, and a combination of the two.<br>We perform a series of simulation-based experiments in which a group of e-puck-like robots must perform an integrated navigation and obstacle avoidance task in environments of different complexity.<br>We show that multiplicative controllers and hybrid controllers maintain stable performance levels across tasks of different complexity.<br>We show that: (i)~summing controllers evolve diverse behaviours that vary qualitatively during task-execution, and that (ii)~multiplicative controllers lead to less diverse and more static behaviours that are maintained despite environmental changes.<br>Complementary, hybrid controllers exhibit both behavioural characteristics, and display superior generalisation capabilities in simple and complex tasks. <strong>Description of the dataset:</strong> The dataset contains a number of figures and .gv files that represent artificial neural networks used as evolved robotic controllers. For a given experimental setup (see the paper), each controller evolved is named as "S.R", where S represents the sample/run in which the controller was evolved, and R is the id of the robot. Networks are described in the following manner: - Ri denotes sensor i for robot detection, with i in [1:8] - Wi denotes sensor i for wall/obstacle detection, with i in [1:8] - E denotes the virtual energy level sensor - Hx denotes the hidden neuron H with id x - LW represents the output neuron that controls the robot's left wheel - RW represents the output neuron that controls the robot's right wheel