Microsoft’s AI Mocks Humans by Notching Up Perfect Score in Ms. Pac-Man

Microsoft's AI Mocks Humans by Notching Up Perfect Score in Ms. Pac-Man

If there was ever a doubt that artificial intelligence could one day overtake humans, then what follows should be sign enough that the day isn’t too far away. Microsoft’s AI has managed to hit the one million mark in cult-classic game Ms. Pac-Man, something humans never managed to achieved in the 35 years of the game’s existence.

Maluuba, a deep learning startup that was acquired by Microsoft earlier this year, created an AI system that learned the ways of the game to reach the coveted score of 1 million points by level 201. So, not only does Microsoft’s AI manage to achieve the highest possible score in both human and AI history, it managed to do so before hitting the all-too well known level 256 glitch. In a video posted by Microsoft Research, you can see the AI reach the game’s maximum point value of 999,990 on Atari 2600, after which the game seems to start over.

It will also interest you to know that the highest recorded score in Ms. Pac-Man – by a human, of course – is 266,330 points, as recorded by HighScore.com. This makes Microsoft’s AI look truly remarkable and gives us a glimpse at how machine learning has evolved over time.

Maluuba was able to set the new Ms. Pac-Man record using machine learning and breaking up the game into small problems with “a separate reinforcement learning agent for each problem,” which the team calls Hybrid Reward Architecture. Here, individual agents are rewarded based on their assigned task. Through this ‘divide and conquer’ method, the top agent gets a feedback from the little agents to understand which is the best route for Ms. Pac-Man to take to avoid being eaten by ghosts, the video explains.

Recently, Google’s AI AlphaGo once again defeated a human, Ki Jie, at the ancient game Chinese game, Go.

Games like Ms. Pac-Man and Go, though quite old, are revered for their complex gameplay, which is why companies in the field of AI test out their machine learning algorithms on them.Maluuba sees an expansive, practical real-world applications though the Hybrid Reward Architecture used in the game, like helping a company predict which potential customers to target, or advancements in natural language processing.

 

 

 

[“source-gadgets.ndtv”]