Reinforcement Learning and Neuroevolution in Flappy Bird Game (original) (raw)

Pattern Recognition and Image Analysis, 2019

Abstract

Games have been used as an effective way to measure the advancement of artificial intelligence. Chess, Atari2600 and Go, are some of the most mediatic demonstrations where AI computer programs defeated human players. In this paper we add the popular Flappy Bird game in the list of games to quantify the performance of an AI player. Based on Q-Reinforcement Learning and Neuroevolution (neural network fitted by genetic algorithm), artificial agents were trained to take the most favorable action at each game instant. The Neuroevolution agent outperformed by far the Reinforcement Learning agent (111 points average result) and achieved on average super-human performance of impressive score of 28700 points.

Petia Georgieva hasn't uploaded this paper.

Let Petia know you want this paper to be uploaded.

Ask for this paper to be uploaded.