Atari is a new AI which, as the name suggest, is related to arcade games. The creators claim that is is 10 times faster than Google DeepMind AI. What sets it apart is the capability of solving problems in environments where actions to achieve a goal are not immediately obvious.
It uses a reinforcement learning technique, which relies on increasing the number of points when performing actions that bring the agent closer towards the goal, and decreasing the points otherwise. As an example, it rewards actions such as ‘climb the ladder’ or ‘jump over that pit’.
Associate Professor Fabio Zambetta from RMIT University had said “We’ve shown that the right kind of algorithms can improve results using a smarter approach rather than purely brute forcing a problem end-to-end on very powerful computers. Our results show how much closer we’re getting to autonomous AI and could be a key line of inquiry if we want to keep making substantial progress in this field.”
However, their findings are not new in the domain, as other similar methods already exist. For example, Unity’s ML Agents use reinforcement learning for multiple types of games, with various tasks, such as: balancing objects on flat surfaces, playing tennis, controlling double-jointed arms, learning to walk, and navigating in a labyrinth to complete a task. The method is well-documented and can be found on GitHub.
Details about how Atari achieves its performance, in terms of technology and algorithms have not been made public yet. An oral presentation will take place at the 33rd AAAI Conference on Artificial Intelligence in Honolulu, which will hopefully shed light on these implementation issues.
Source (Atari master: New AI smashes Google DeepMind in video game challenge, RMIT Australia, 31.01.2019)
Original paper: Dann, M., Zambetta, F. and Thangarajah, J., 2019, July. Deriving subgoals autonomously to accelerate learning in sparse reward domains. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 33, No. 01, pp. 881-889).
