OVERVIEW
MANAGEMENT
PERFORMANCE
POSSIBILITIES
CAPITALS
ACTIVITIES
ACTORS
BURGESS
Technology | ||
|
Burgess COMMENTARY Peter Burgess | ||
|
Original article:
AI news Pamela Faulkner Mar 4, 2021, 4:14 PM (14 hours ago) to Jerome, me Yet another major breakthrough in the world of AI… Artificial intelligence (AI) has just taken another step forward. Researchers from OpenAI and UberAI Labs have developed an AI that is able to master classic Atari games from the 1980s. Believe it or not, this is a big milestone. The research team unleashed the AI on old video games like Centipede, Berzerk, Pitfall, and Montezuma’s Revenge. These are all games that I played as a kid in the ’80s. I’m sure many readers will remember them as well. While these games look simple and basic by today’s standards, they are deceptively difficult for an AI. That’s partly because advancing in the games isn’t a linear process. Players must solve puzzles, which often require some backtracking to previously explored areas in the game. For this reason, even Google’s DeepMind couldn’t produce an AI capable of mastering these old games. DeepMind’s Agent57 AI came close, but it had some flaws that held it back. The main problem with Agent57 is that it operated using what’s called intrinsic motivation. This is a model where the AI is rewarded for discovering new aspects of a game. But this approach also leads to what AI researchers call “detachment.” Because the AI is rewarded for learning how to advance to new areas, it sometimes forgets to go back and explore old areas later in the game. It becomes “detached” from those parts of the game. OpenAI and UberAI Labs solved this problem. And their solution was elegant in its simplicity. The team programmed their AI to “archive” areas of the game as it explores them. The AI makes note of certain areas that may be useful later. Then it revisits these areas later in the game. It comes back to see if there’s something new that can help it advance in the game. And sure enough, this approach helped the AI master these old Atari games. It can now beat these games faster than human experts. I know this may sound like a silly application for AI, but this is a big deal. Now that this AI has proven its ability to archive old information to check later, this same approach can be turned loose on other areas of research. This could be applied to supply chain management, materials design, therapeutic development, and many other more practical areas of research. I can’t wait to see this new tech applied to tough optimization problems. I expect we will see more breakthroughs made using this research in the coming months.
| The text being discussed is available at | and |