MEET Go-Explore


If you’re like me you’ve likely spent countless hours yelling at a video game because you just can’t seem to get it right. Well, I guess researchers at UberAI and OpenAI have felt a similar pain before because they recently unveiled “Go-Explore,” a new family of algorithms that have been successfully completing the most notoriously difficult Atari Games that even previous AI couldn’t beat. Go-Explore goes back to the roots of what makes a good AI, which is essentially the process of trying something, failing, and trying it again with a modification until eventually you have a perfect result. By utilizing this reward system of positive reinforcement and combining that with intelligent exploration, Go-Explore has reignited the classic development of AI and has done so in a fun way that reminds me it’s okay to spend hours on hours repeating the same level so long as it leads to eventual success.