Title -------------------------------------------------------------------------------------------------- Topic Path Planning or Robot Navigation where usually: From Start to Goal minimizing its route -------------------------------------------------------------------------------------------------- Not so difficult A star Genetic Algorithm Reinforcement Learning Neural Network Ant colony -------------------------------------------------------------------------------------------------- Example by GA -------------------------------------------------------------------------------------------------- Extention: Maximizing exploration -------------------------------------------------------------------------------------------------- ... -------------------------------------------------------------------------------------------------- From Camel to Jeep Further to Planet Landrover -------------------------------------------------------------------------------------------------- Further Variation Mars Landrover From A (start) to B (goal) minimizing its route => From A to A maximizing its route? -------------------------------------------------------------------------------------------------- Still not so difficult if we use a huristics example -------------------------------------------------------------------------------------------------- What is intelligence? Pardon? repeat intelligence should be spontaneous, flexible, or unpredictable more or less A-star, GA, performance were intelligent? -------------------------------------------------------------------------------------------------- Maybe Reinforcement Learning under "epsilon greedy strategy" would be an option -------------------------------------------------------------------------------------------------- But in a world of no obstacle, no wall no corridor only goal state is given a reword -------------------------------------------------------------------------------------------------- Our aim spiking neuron who learn during random exploration -------------------------------------------------------------------------------------------------- D. Meunier and H. Paugam-Moisy (2005) "Evolutionary supervision of a dynamical neural network allows learning with on-going weights." Proceedings of International Joint Conference on Neural Networks (IJCNN) pp. 1493-1498 Up to now, nobody (to the best of our knowledge) has been able to show how it is possible to learn with STDP, although there are growing evidence that such a mechanism is quite general in the nervous system [9]. or M. A. Farries and A. L. Fairhall (2007) Reinforcement Learning With Modulated Spike Timing. Dependent Synaptic Plasticity "Although synaptic plasticity is widely believed to be a major component of learning, it is unclear how STDP itself could serve as a mechanism for general purpose learning." -------------------------------------------------------------------------------------------------- F(\Delta t) = A_{+}\exp(\Delta t / \tau_{+}) if \Delta t < 0 F(\Delta t) = - A_{-}\exp(- \Delta t / \tau_{-}) if \Delta t > 0 \tau_{+} = \tau_{+} = 20 ms A_{+} = 0.005 -------------------------------------------------------------------------------------------------- eps aka echo state network resevoir computing liquid state.. nevertheless Extention of the problem Jeep & Landrover Modular-neural-network-and-classical-reinforcement-reinforcement-learning-for-autonomous-robot-navigation:-Inhibiting-undesirable-behaviors