AI Learns to Play Pokemon Red

Multiplayer titles might not need human players in the future because AI keeps learning new tricks every day, including the rules of various popular games. The recent advancement in this field comes from Peter Whidden, who managed to train AI to play Pokemon Red using reinforcement learning.

Reinforcement learning offers the agent (AI) rewards for completing tasks correctly. With this, it learns what to do in order to maximize the reward. As Whidden explains, the agent takes in images from the screen and chooses which buttons to press to optimize its choices. You can see it in action in this hilarious video by AI Warehouse where an AI named Albert teaches himself to walk.

But let's go back to Pokemon. The video shows how AI was gradually learning the game, and it's a long process. Interestingly, when it figured out how to get out of the starting area, it got distracted, just like we, humans, do. It turned out that the algorithms made it so chilling outside and admiring the view was more rewarding than exploring, so the rules had to be changed.

Eventually, the agent adjusted, fought its many battles, and caught its own Pokemon collection, so the experiment was a success, although it could be improved in the future.

This is not the first video game for AI. NVIDIA has taught its LLM-powered embodied lifelong learning agent Voyager to play Minecraft in context using GPT-4. It's fascinating, take a look.

And if you're interested in reinforcement learning, check out Ubisoft's approach to NPCs and Google DeepMind’s paper that discusses a method of training AI to play “infamously hard exploration games” using YouTube videos of human playthroughs.

Also, join our 80 Level Talent platform and our Telegram channel, follow us on Instagram, Twitter, and LinkedIn, where we share breakdowns, the latest news, awesome artworks, and more.