As a recent convert to the world of Pokémon, thanks to my 8-year-old’s enthusiasm for the card game, I’ve found myself reflecting on the nostalgia of the original GameBoy Pokémon games. I remember when these games first came to the U.S. in the late ’90s, when my friends with GameBoys were taking on the role of a Pokémon trainer and aiming to “catch them all” in the original Pokemon Red and Blue.
Fast forward to today, and I came across a fascinating video titled Training AI to Play Pokémon with Reinforcement Learning that really sparked that old memory. The video, made about a year ago, showcases how AI can be trained using reinforcement learning to play the original Pokémon GameBoy game. The concept of reinforcement learning is particularly intriguing because it’s based on training AI systems to make decisions based creating “a gentile curriculum of rewards” and penalties, much like how humans learn through trial and error.
In the video, the AI is tasked with navigating the Pokémon game environment, making strategic decisions to progress in the game, battle trainers, and, of course, catching Pokémon. The AI essentially learns the best strategies for success over time, all based on the rewards it has learned. What’s particularly interesting is that the AI not only learns how to play the game, but shows how gameplay and game design impact learning. For instance, there was a point where it found reward in just being in a certian area, rather than progressing though the game.
This approach to AI exploration is a fantastic example of how machine learning can be applied in unconventional, and nostalgic, ways. It also sparked a flashback to a computer science student I had back in the STC days, who was exploring an early vision machine learning model to automate Pokémon gameplay to fill up their Pokedex faster.