Eventos Anais de eventos
COBEM 2019
25th International Congress of Mechanical Engineering
Neuroevolution approach for reinforcement learning task applied to control system of a cart pole system
Submission Author:
Leandro dos Santos Coelho , PR , Brazil
Co-Authors:
Marco Boaretto, Gabriel Chaves Becchi, Luiza Scapinello Aquino, Aderson Cleber Pifer, Helon Vicente Hultmann Ayala, Leandro dos Santos Coelho
Presenter: Helon Vicente Hultmann Ayala
doi://10.26678/ABCM.COBEM2019.COB2019-1044
Abstract
Designing optimization methods for artificial neural networks (ANNs) and deep learning is very challenging due to the non-convex nature of the optimization problems. The tuning of the control hyperparameters can be an exhaustive and time-consuming task given the number of parameters that need to be tuned, such as the number of hidden layers, number of neurons in the hidden layers, activation function, optimization method, the percentage of neurons that will suffer dropout require skills of the designer. Evolutionary algorithms and swarm intelligence paradigms, classes of metaheuristic approaches, are well suited to address this challenge of tuning of the control hyperparameter. The contribution of this paper is compared the performance of three stochastic optimization metaheuristics including genetic algorithm, particle swarm optimization and grey wolf optimizer to find the best architecture of a feedforward ANN model provided a limited search space of control hyperparameters. The ANN performed a reinforcement learning task, which consists of the control of a cart pole in a virtual environment simulation from the OpenAI gym toolkit. The obtained results were promising in terms of the optimized ANN model performance to control the cart pole system.
Keywords
Control Systems, machine learning, Deep learning, Evolutionary algorithms, Reinforcement Learning
