A great look at the history of machine learning (it started with matchboxes in ’40s). The first machine learning setup (it wasn’t a computer) was designed to play tic-tac-toe. It shows that machines don’t learn in the way humans do; we go for 3 in a row, the machine picks moves based on the present board without understanding the concept of “3 in a row”. It is important to understand the parameters of the problem you are trying to solve, you need to run the program/algorithm/robot a whole bunch of times so it can learn, and you should vary the learning sets if possible. Or make them really, really large. You don’t want to train the machine on a faulty data set because then it will only learn faulty knowledge. For example, training it against a player that makes the optimal move every time with teach the machine to play to a draw because winning is impossible. You also need to determine how to give credit to different elements in a solution chain (like GA attribution models).
Takeaway: Use lots of good data and spend as much time as you possible can thinking through the problem up front so you can map the system properly.
Skynet is far off update: AlphaGo would have melted down had the board size they played on been anything other than 18×18