Microsoft AI Beats Ms Pac-Man Top Score

The AI training techniques could be used to train better AIs for other tasks

A Microsoft  artificial intelligence (AI) system has managed to achieve the perfect score of 999,990 in retro arcade game Ms Pac-Man. 

A team from deep learning startup Maluuba,  which is owned by Microsoft, used a form of reinforcement learning,  a form of learning through rewarding or punishing actions carried out through trial and error. to train an AI to play the perfect game of the 1980s video game for the Atari 2600. 

Ms Pac-Man AI

Beating Ms Pac-Man 

To achieve the highest ever score in Ms Pac-Man, the team decided the challenge of mastering the game into segments to be distributed amongst AI agents rather than relying on a single AI to figure out the game by itself. 

The method, dubbed Hybrid Reward Architecture, put 150 AI agents to use, were each one works in parallel to master parts of the game, such as collecting the dots in the arena or avoiding the enemy ‘ghosts’. 

The researchers then selected the top performing AI agent and used it as a for of managing agent whereby it took a balance of suggestions from all the other agents as to where beat to move the player character Ms Pac-Man. 

For example, if 100 agents wanted Ms Pac-man to move right to collect a pellet, but three agents wanted to move left to avoid confronting an enemy ghost, then the master AI would prioritise the latter move over the former.  

“This idea of having them work on different pieces to achieve a common goal is very interesting,” said Doina Precup, an associate professor of computer science at McGill University in Montreal, who noted that the AI system is a notable achievement as it beat a game known to be notoriously difficult to crack. 

Having created an AI system to beat Ms Pac-Man, the techniques used to train it could be used to better train AI agents to perform tasks to aid people in their working and daily lives. 

“That would be really, really exciting because it’s another step toward more general intelligence,” said Precup.

The idea of training  AIs with games comes form the human-level interactions they promote. However, there is still some debate as to whether AIs can replicate these behaviours for other tasks or if they can only be trained for specific games and applications. 

Microsoft has had mixed results with AI so far; its machine learning powered virtual assistant Cortana has found relative success and is being increasingly embedded across Redmond’s software portfolio.

But Microsoft also created a Twitter chatbot called Tay designed to mimic the speech patterns of a 19-year-old American girl, but Tay was quickly canned after it started spouting offensive and racist tweets derived from public data on Twitter, a place not know for the most logical and polite material. 

As such, perhaps people should worry less about machines stealing jobs and more about them showering people with a barrage of abuse. 

Put your knowledge of artificial intelligence to the test. Try our quiz!