Now AI can help you to “catch ’em all”!
In an awe-inspiring feat of technology and gaming fusion, a skilled Pokémon Red player has leveraged the potent capabilities of Reinforcement Learning to teach artificial intelligence (AI) to conquer the captivating world of Pokémon. In a digital age marked by constant advancement, the convergence of AI and video games is a natural progression. AI has already showcased its creative talents in art and music, and now it embarks on the quest to become a gaming maestro.
Compilation of a range of distinct gameplay sequences the computer is experiencing. (Screenshot from Peter Whidden’s Youtube video)
Enter Peter Whidden, a renowned YouTuber who has meticulously documented this extraordinary journey in a captivating 33-minute video. This experiment distinguishes itself through its staggering scale, with the AI playing over 20,000 games of Pokémon Red over five simulated years.
Initially, the AI commences its Pokémon Red adventure with a phase of seemingly random button mashing and aimless wandering. However, as the simulated game time unfolds, the AI progressively accumulates knowledge about effective strategies and optimal gameplay.
This video provides a succinct encapsulation of the AI’s evolution, highlighting,
“It starts with no knowledge whatsoever, and is only capable of pressing random buttons. But, throughout five years of simulated game time, it gains many capabilities by learning through its experiences.”
A pivotal juncture arises when the AI grasps the fundamental game controls and navigates the optimal paths to progress in the Pokémon game.
As time unfolds, the AI’s proficiency in Pokémon Red becomes increasingly apparent, as it embarks on a journey through diverse gameplay loops. Notably, the AI derives its skills from an extensive dataset of past experiences, often numbering in the tens of thousands. In a remarkable development, the AI even attains the ability to manipulate RNG (Random Number Generator).
Reinforcement Learning plays a central role in nurturing the AI’s development, enabling it to learn from high-level feedback provided by the player. Peter Whidden delves into the intricacies of setting objectives for the machine based on a point system. For instance, the AI prioritises conquering Gym Leaders over aimless wandering, ensuring efficient progress within the realm of Pokémon Red.
Screenshot from Peter Whidden’s Youtube video
For those intrigued by the idea of embarking on this AI journey, it is indeed possible to implement this model on your computer. The comprehensive “Pokémon Red Experiment” GitHub repository contains all the essential information for this undertaking. Be aware, though, that this experiment demands substantial computational resources, consuming nearly 100 GB of RAM by default. Nevertheless, it represents a remarkable milestone in the ever-evolving landscape of AI technology.
Even if you choose not to personally dive into this experiment, I strongly recommend immersing yourself in Peter Whidden’s complete AI Reinforcement Training video. It truly offers an unmissable experience, providing an in-depth insight into this extraordinary technological achievement where the worlds of AI and gaming beautifully intertwine, weaving a captivating narrative of skill acquisition, learning, and mastery.