Developer of Openai and Ober have presented their KI system called Go-Explore in the "Nature" journal, the Atari game classist such as Pitfall or Montezuma’s Revenge better than man dominated. Previous KI systems have been difficult with the appropriate type of play, as they lacked a clear indication of the evaluation of a randomly executed train in the procedures of "Reinforcement Learning" they used. Go-Explore does not explore the way through the games levels after the pure random principle, but checks whether the procedure has guided really close to the destination.
At the Atari classic Montezuma’s Revenge, the Ki succeeded in beating the existing world record, with pitfalls overrise their performance at least that of an average player. The researchers speak of a breakthrough and ame that the involvement of human prior knowledge (Domain Knowledge) could provide, for example, in the field of robotics for progress. The authors have also made Go-Explore accessible via Github.
70 years old dream
Prof. Jan Peters of the TU Darmstadt sees a change in time in the KI, if GO-Explores should be achievable.
Jan Peters, Professor for intelligent autonomous systems at the TU Darmstadt, looks different: "If we have good simulators and can clearly define problem as well as situation, a problem in robotics is usually easy to relax." he explains opposite the Science Media Center.
Nevertheless, for him is a rough surprise in the details of the nature article: "If Domain Knowledge is included, Go Explore can beat the human world record. If this statement turns out to be generalized, then this can be a change of time in the Ki, where the two Ki families of the statistical-neuronal procedures and the domain knowledge engineering finally unite. For many Ki researchers this was a nearly 70-year-old dream!"