AI from IBM learns by playing Flappy Bird

IBM scientists have unveiled research on how machines can constantly learn tasks, getting better and better instead of getting stuck on one level. The game Flappy Bird was used as a test, writes ZDNet.

The scientists carried out research into something that is also known as ‘lifelong learning’ or ‘continuous learning’. That area has been explored for decades, but remains a challenge. time.

Flappy Bird was for IBM’s scientists the most important test for their research. The game, which was removed from smartphones in 2014 by maker Dong Nguyen because it was too addictive, lets users fly a small bird along all kinds of pillars. The intention is not to hit the poles.

The IBM researchers defined every change in the game – such as the height of the poles – as a new task. Neural networks must then extrapolate what they have learned in one task to another, in order to be able to make maximum use of what they have learned in the past. This is also called Meta-experience replay (MER).

Tests

This approach was tested with two different benchmark tests for neural networks. One is a version of the traditional ‘MNIST’ handwritten data set developed by the National Institute of Standards and Technology. The aim is to identify labeled examples of numbers written on different shapes and rotations.

The second test was with Flappy Bird, using a reinforcement learning approach based on a neural network called Deep Q Network (DQN). In both cases, the researchers state that the accuracy scores are better than in the benchmarks.

The DQN that MER used becomes “a platinum player at the first task, while learning the third task” in Flappy Bird. “DQN-MER shows the kind of learning patterns that people expect from these games, while a standard DQN has difficulty in generalising changes in the game and in retaining knowledge about time,” said the scientists.