Ms. Pac-Man is best known for its unpredictability, which makes the Microsoft AI's efforts all the more special as no one has done it before.
In what is being seen as further confirmation of the relative superiority of well-designed AI systems vis-a-vis their human counterparts, a Microsoft subsidiary, Maluuba that specializes in deep learning techniques has managed to reach the magical Ms. Pac-Man score of 999,990 on Atari 2600.
That makes it four times better than the highest known score of any human being at the game that is best known for its unpredictability. The best a human has ever managed on the Atari 2600 version is 266,330.
Maluuba is the Canadian deep learning start-up that Microsoft had acquired early this year. Also, what makes the Maluuba led effort even more impressive is the strategy adopted to tackle the unique challenges that the 80s classic arcade games offer.
As sources revealed, the Maluuba team adopted a divide and conquer approach to tackle the game. So instead of using a single agent to learn all the tricks of the game and proceed, around 150 agents were deployed with each having a relatively simple and specific role to play.
Further, a reinforcement learning strategy referred to as ‘Hybrid Reward Architecture’ has also been used to deal with the performance of each agent. So any agent that managed to eat the pellet went in for rewards while the others tired running away from the ghosts.
The performance of each of the agent were then analyzed by a top agent, the role of whom is likened to that of a senior manager in any organization. The top agent weighed in the performance of each agent to eventually decide on relocating Ms. Pac-Man.
While all of this makes for a stupendous scientific feat, the real world implication of the above development is all too obvious as well. For experts believe a similar approach will also help such AI systems to arrive at a decision while dealing with unpredictable real life situations. So whether it is up to decide if you need to buy or sell in the stock market or predicting company’s sales performance, such systems can find extensive applications in future.