Learning When to Kick through Deep Neural Networks

• Dicksiano Melo ITA
• Carlos Forster ITA
• Marcos Máximo ITA

Resumo

This paper propose a novel approach to solve a Decision Making problem in RoboCup Soccer 3D Simulation League environment. Instead of using heuristics, we utilize a Neural Network that is trained from a large number of events, in order to achieve a reliable approach for a recurrent Decision Making problem: given a virtual agent which dominates the ball, decide if it is capable to kick the ball, given the state of the soccer field. In order to achieve the best from the data collected, which is certainly the bottleneck, we utilize Neural Architecture Search through Genetic Algorithms.

Palavras-chave: Neurons, Robots, Neural networks, Meters, Three-dimensional displays, Games