John Boyd was a jet fighter pilot. He wanted to create an intuitive system that could be used in war scenario where two fighter jets try to intercept one another. He believed that the most interesting and agile learning process in this situation would be to observe the scene, orient our minds, decide on an action, then act to see if the action has desired outcome or not, and run this loop again and again.
Boyd argued that, as long as you iterate this loop and despite any mistakes made along the way you will eventually win. You can shoot a missile that doesn’t hit the target or embark on a project that can’t be implemented, but as long as you learn from your mistakes and your mind remains agile, you will eventually win by iterating the loop. This is a great analogy for machine Intelligence ,since all machine- intelligence models operate like OODA loops.
Machine-Intelligence models observe the data we have and orient around an initial point. That is, they choose a random starting point in a given data space and then make moves based on their mathematical relations to the data. beginning at and moving beyond the chosen starting point. This process yields information that can be acted upon.
Think of a robot that tries to flip pancakes. Before it begins to act, it has been fed data on how to flip pancakes. The pancakes-flipping robot hand has a certain range of movement, can apply more force or less, and so on.
The first thousand tries, the robot can’t flip the pancake. It’s either too fast or too forceful, and the half cooked pancake falls out of the pan. The robot makes a lot of mistake along the way but eventually learns from experience, iterating the OODA loop. Eventually, it perfects all the parameters of efficient flipping. The robot now flips pancakes perfectly, without burning or dropping them on the floor.
It takes our brain time to learn. Think about how long it took you to learn to use a knife and fork. We eventually accomplish the task because we learn how to think about and solve problems from experience over time.
It’s the same with machine learning. Although machine learning tends to be extraordinary agile, it still takes time. In the orient phase, the machine at first chooses a random point in the data space, but eventually over time, this random point will converge to what data scientists call a global minima, the most efficient and optimal point in space. When you reach a global minima, you have found the best way to solve a problem.
The machine will make mistakes while learning, like humans do, but also like humans most of the time the machine will ultimately be able to perfect the task.