Imitation Learning

What is Imitation Learning?

Imitation is self-explanatory in definition; simply put, it is the observation of an action and then repeating it. So far, this is an inherently “living” concept, and one that is difficult to reproduce in AI.

A Practical Example in Artificial Intelligence

AlphaGo, a computer that beat Lee Sedol at his own mastered game, did not learn to play Go by imitating others. Instead, AlphaGo played Go thousands upon thousands of times and kept track of its results every time it played. It kept a record of the policies that worked well, and which didn’t, and let the positive policies influence its next game. Good results were reinforced, but AlphaGo did not master Go by watching Lee Sedol beat his opponents. If it had, that would have been imitation learning. 

Why Imitation Learning is Difficult for AI to Master

There are some problems that make reinforcement learning difficult for AI. While it is somewhat easy to measure the concept of success in some something simple like a game of Go, success is not always that straightforward, and therefore not always easy to obtain.

In addition, reward sparsity is an issue with computers. This is where what is considered success is rare, and the computer cannot learn from its mistakes because it does not know which are bringing it closer to a “success”.