Implementasi Q-Learning dan Backpropagation pada Agen yang Memainkan Permainan Flappy Bird

Ardiansyah (Universitas Komputer Indonesia)
Ednawati Rainarli (Universitas Komputer Indonesia)

Jurnal Nasional Teknik Elektro dan Teknologi Informasi
Vol 6 No 1: Februari 2017

This paper shows how to implement a combination of Q-learning and backpropagation on the case of agent learning to play Flappy Bird game. Q-learning and backpropagation are combined to predict the value-function of each action, or called value-function approximation. The value-function approximation is used to reduce learning time and to reduce weights stored in memory. Previous studies using only regular reinforcement learning took longer time and more amount of weights stored in memory. The artificial neural network architecture (ANN) used in this study is an ANN for each action. The results show that combining Q-learning and backpropagation can reduce agent’s learning time to play Flappy Bird up to 92% and reduce the weights stored in memory up to 94%, compared to regular Q-learning only. Although the learning time and the weights stored are reduced, Q-learning combined with backpropagation have the same ability as regular Q-learning to play Flappy Bird game.