AI can now learn to play Pokemon Red after 5 years of simulated game time

One brilliant Pokemon Red player has trained AI to play the game through Reinforcement Learning. Technology is always growing, so it's not a surprise to see this happen eventually. AI is already capable of creating surprisingly detailed artwork and music. Let's now see how it fares when playing video games.

In this case, YouTuber Peter Whidden released a video documenting his journey. It's over 33 minutes long, yet it's a fascinating watch for anybody curious about machine learning and similar fields. Over 20,000 games of Pokemon Red were played by an AI, with five years of simulated game time involved.

Needless to say, it's intriguing that somebody trained a machine to play a fairly complicated video game from start to finish.

---

How one person trained AI to play Pokemon Red through Reinforcement Learning

The AI starts off Pokemon Red by mashing random buttons and wandering aimlessly. Due to how long it has been playing the game, it eventually learns what works and what doesn't.

This video states the following regarding the AI:

"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."

There comes a point when the machine learns the basic controls and the routes it should take to advance through the Pokemon game.

---

The AI gets better in Pokemon Red with more time

All these screens show off various different gameplay loops that the computer is going through (Image via Peter Whidden)

Remember, the AI is essentially getting data on what worked for past experiences, which can scale up to the tens of thousands. It gets so advanced that this program even learns how to manipulate RNG.

Reinforcement Learning allows the AI to learn if the player gives it high-level feedback.

Peter Whidden goes into more detail regarding how to set up the objectives for the machine to aim for based on a point system. For example, beating Gym Leaders is a bigger incentive for the AI than just walking around. Over time, the AI knows what is worth doing in Pokemon Red, leading it to make more progress than how it fared on its early playthrough attempts.

---

Players can run the AI code themselves

Another screen shows off how often this AI played the game (Image via Peter Whidden)

It is possible to run this model on one's own computer. Here is a link to the GitHub that has all the relevant information you'll want to carefully read:

• Pokemon Red Experiment

Get Python 3.10 if you don't already have it, and you can get started on training a model, although this will take up nearly 100 GB of RAM by default. Nonetheless, this sort of experiment is a marvelous development in AI technology.

Even if one doesn't want to run this program on their own computer, it's still worth watching the full AI Reinforcement Training video by Peter Whidden since a simple article doesn't do it enough justice on its own.