Your AI Is Learning by Playing Computer Games!
In order for AI to learn a task, it needs data. Masses of data. As Rene Buest, Director of Technology Research at Arago, refers to Prof. McCarthy’s definition in his article “AI is about machine reasoning – or when machine learning is just a fancy plugin”, AI should be considered as a vigorous system that needs unstructured input to train its senses and along with it a semantic understanding of its surroundings in order to (re)act accordingly.
But where do I get this input – this data that makes my AI learn and develop?
If for example you’re building a self-driving vehicle and collecting data meant taking the car onto the streets – this would be laborious and time-consuming. Indeed, gathering valuable data requires a lot of time, deep pockets and access to lots of resources. But what if you want to take on the likes of Google and Tesla but don’t have their financial means?
Here is an idea and approach taken by Adrien Gaidon, a Research Scientist in the computer vision group at Xerox Research Centre Europe:
In order to meet the need for massive sets of training data, his team used the Unity videogame engine to develop an exceptionally realistic environment full of everyday objects and surroundings, such as sidewalks and motorbikes, to see if a deep-learning AI could learn from this virtual space. The potential and benefit of this approach is on the one hand to e.g. have the AI run through specific scenes over and over again in order to improve its skills but on the other hand to have the ability to change environments and create new scenarios from which the AI can learn.
South Korean startup Mars is also looking to develop an AI for use in trucks. They were searching for efficient methods to train their AI and also turned to the world of gaming, selecting Euro Truck Simulator 2 as their testing ground. The game attempts to accurately replicate the day-to-day activities of a truck driver and simulate a realistic environment which has, as a matter of fact, proven highly effective for the team. Data and insights that were generated through new situations enabled their AI to advance its skills and learn about a wide range of road hazards.
All in all, the gaming industry has proven to be a great testing and training ground for AIs, generating vast amounts of valuable data that can be applied to autonomous systems.