Modelo de rede neural de múltiplas escalas de tempo aplicado ao robô Dimitri

Abstract

The idea for the implementation and application of the Multiple Timescale Recurrent Network (MTRNN) model started with experiments based on machine learning (Peniak et al., 2011; Nobuta et al., 2011). In which a humanoid robot learns complex activities through interaction with humans. The MTRNN model (Yamashita and Tani, 2008) is distinguished by having two main types of units, the fast neurons and the slow neurons. This characteristic of the network makes it able to encode and reproduce in detail dynamic phenomena in fast neurons and at the same time represent these phenomena in a higher degree of abstraction in the slower neurons, bringing an idea analogous to the intentionality of the actions observed in humans and in others Animals. The main goal of this work is the application of the MTRNN model in the Dimitri robot, recently developed at the UFSM Technology Center. The neural network was implemented in Python, and its implementation started in a previous work. The network training uses the Backpropagation Through Time (BPTT) learning algorithm optimized by Nesterov's accelerated gradient method. The data used for the training of the network refer to the position of the servo motors when being manipulated by a tutor, who will perform cyclical movements. Experiments were carried out to test the network and its characteristics. The implemented network was able to learn three different movement patterns, but was unable to operate in closed loop due to sensitivity to disturbances, so the network is not ready to be applied directly to the Dimitri robot.