Achieving Human Expert Level Time Performance for Atari Games – A Causal Learning Approach
Seng-Beng Ho, Xiwen Yang, Therese Quieta · 2020
In a previous paper, we have argued that what constitutes human-level performance for Atari games must include the measurements of both the score achieved as well as the time taken to learn to achieve it. When both of these measurements approach the level of humans, the method and system will also become practical as high score and short time performance for learning can lead to practical real-world applications. Other well-known approaches, such as DeepMind's Atari game player, achieve human-level scores but the time taken to learn to play the games is many orders of magnitudes slower than that of humans, leading to inapplicability in real-world scenarios beyond the games. In our previous paper, we demonstrated the ability of a causal learning framework to learn to play the Atari game, Space Invaders, at both the score and time performance levels of a human novice. In this paper, we show that an extension of this system reported earlier leads to the ability for the system to perform at both the score and time level of a human expert. A major innovation of this paper is a “changing goal tracking” system that is not only applicable to Atari games such as Space Invaders, it is also applicable to general AI systems.