Robots and computers are now an integral part of today’s society, automating tasks such as manufacturing items, parking cars, and performing surgeries. However, most robots are only as good as the programmers who write their instructions. However, what happens when a robot is faced with a situation that it has not been programmed to perform? To solve this problem, programmers are utilizing the concepts of Artificial Intelligence (AI) and machine learning (ML) to prepare robots for every new challenge. Many techniques have already been developed in ML but they have to be specifically tailored to each task and may take thousands of “training examples” for the computer to “learn” the correct process. To improve one of the methods, we wrote a simple, turn-based board game and two computer opponents to go with it. One of the opponents was designed using artificial intelligence, where the computer evaluates each move against a set strategy that we supplied it with. The other was designed using machine learning, where the computer evaluated each move against a strategy that it created. The two players were pit against each other hundreds of times so that the ML opponent could refine its outcomes to the perfect strategy. The ML opponent used the score difference of each game to evaluate the strategy it used and refine it over time.
After 1000 games, the strategy of the ML player showed that being offensive was about 3 to 5 times more important than being defensive but also that sabotaging the opponent is only equally as important as furthering your objective. Although the data shows a general convergence toward this strategy, the amount that the improving ML opponent won against the static AI opponent rarely fluctuated. The ML opponent was consistently superior, but it lost at a similar rate across the entire experiment, meaning that the ML opponent might not have actually significantly improved, as was the objective.
Although it may not improved, it was still more effective than the static opponent, proving that it is possible for machine learning-programmed software to exceed that of hand-programmed software. You likely see this all around you, where the perfect marketing strategies and search results have been optimized by none other than machine learning.