Did someone say Skynet? It turns out that a group of scientists is working, as part of the Uber AI Labs team, on teach an artificial intelligence to complete platform games from the 80s, What Pitfall.
The team has created a series of algorithms that can create files of the areas already explored, allowing artificial intelligence to remember the levels. This can then be used in the exploration of new unexplored ones.
This method differs from the artificial intelligence that we have seen so far; Until now, the motivation was to go to new areas, something inappropriate because it can cause them to forget places they have already been.
Using this new method, the team has been able to teach artificial intelligence how to beat games like Montezuma’s Revenge, Freeway and Pitfall; all pretty challenging classic games from the Atari era.
“Our method is actually very simple and has one direction, as is often the case in science,” said scientists Adrien Ecoffet, Joost Hizinga and Jeff Clune in an interview with the BBC.
This new design, known as Go-Explore, is capable of separating the process of exploring new areas and returning to old ones, using different methods to achieve both objectives.
As is often the case in science, this research is being conducted with a much broader goal than teaching a computer how to play. The team says the algorithms could be used to guide robots between industrial locations, or even for dependent robots in a coffee shop of the future.
“In addition, with Go-Explore, we have worked on experimental research in language learning, since they are able to learn words from a text game, or we can even analyze the failures of self-driving cars”, added the team.
