Last year Microsoft shared that its AI has successfully broken the all-time high-score of Ms. Pac-Man. The all-time high score set by humans was 266,330 and Microsoft’s AI achieved a perfect score of 999,900 on Atari 2600.
Microsoft has now used one of the sessions at Build to explain how the company was able to achieve the perfect score. Basically, the company used something called as Reinforcement Learning where software agents ought to take actions in an environment so as to maximize some notion of cumulative reward.
Microsoft said the AI used Reinforcement Learning to break up different portions of the game like collecting pellets or avoiding ghosts into 163 smaller chunks, assigning each task to an “agent.” Each agent then submits its recommendation to a “senior manager,” which makes the ultimate decision on how to move in the game. For example, the senior manager will always choose to avoid a ghost over collecting additional pellets.
Microsoft says that this strategy of aggregating small-scale recommendations to achieve a large-scale goal can be applied to natural language processing and to places like improving Cortana.
If you’re interested in watching the whole session then you can do that on Microsoft Developer’s YouTube channel on click here. Do let us know your thoughts on this in the comments section below.