A programmer trained an artificial intelligence model for 50,000 hours how to play pokemon red, which created an algorithm that was able to explore the game and build a team to defeat the first gym leader, but not one that would find its way through Mount Moon or know that it’s best not to keep buying Magikarp. All in all, this exercise is an interesting way to understand how machine learning works.
Described in an extensive video by Peter Whidden i am a You can interact with the game through normal controls in an emulator. Press a button and watch the screen to see what’s happening like a human player. Each game has a two-hour secret set learning session, although with accelerated simulation, those sessions can be completed in about six minutes of real time; And the process was further accelerated by running 40 test sessions simultaneously.
Since a machine algorithm doesn’t inherently care about beating a video game, Whidden sets specific goals for it. i am a was awarded. To encourage curious exploration, i am a Every time he sees something new he earns a reward point, which is measured by the number of pixels displayed on the screen.
This had some unintended consequences: i am a For example, he looks at the slight animation of water, fascinated. But generally speaking, it motivates the computer to move from Pallet Town to the Green Forest and reach Silver City, where the first gym battle against Brock takes place.
The i am a More rewards and punishments are needed. With rewards focused on seeing new things, i am a He just wants to move on, which means he doesn’t care about fighting or capturing pokemon, so at first I ran away from every encounter. So Whidden added a system where i am a He is awarded based on the total level of his active party pokemon. Work for that i am a will fight for xp and capture pokemonBut this was an unintended consequence.
when i am a I’m going to a Pokemon Centercontact with pc There and some submissions pokemon. This greatly reduced the team’s overall level, suddenly taking away a lot of reward points. It was roughly equivalent to a traumatic experience i am awhich led him to avoid Pokemon Center completely; Until Whidden adjusts the reward system again.
Given that i am a Basically he’s just doing random things until he’s able to discover something that gives him reward points, the fight against Brock becomes a particular problem, as you have to take advantage of his initial weakness. pokemon Rock types can cause them real damage. Only because of a certain iteration of Squirtle’s i am a turned out to be without pp For all moves except bubbles, the algorithm figured out how to beat the gym.
However, though i am a It’s not just about discovering things that might come quite naturally to human players, it’s about learning other, more arcane things very quickly. Whidden at one point realized that the algorithm always traced a very specific and seemingly illogical path from Pallet Town to the first encounter. pokemon Wild This seemed strange until it became clear that this precise sequence of inputs guarantees that pokemon Wild can be captured with a single throw tell me. yes i am a Intuitively learn the same art of RNG manipulation that speed runners develop over the years.
Beating Brock identified a fairly natural end goal for the project, but Wheeden let it go i am a It would take longer to see what would happen, and he went quite far into Moon Mountain, but the monotonous paths of the dungeon were so discouraging. i am a Who never found his way to the other side, so he never reached Celeste City’s Second Gym.
However, some of that i am a Had to buy Magikarp. The shady guy who sells you the worst pokemon All the time at a ridiculous price is practically a joke at this point, but for i am abuy it Magikarp This is a quick way to get five more levels pokemon In your team, the best offer in the game! Apparently, the i am a That Magikarp has bought over 10,000 times.
Oh, and for one last anecdote about the magic of a computer doing random things: At one point, i am a He captured a ratta and named it ‘A.I‘ Sometimes these things turn out to be a little too perfect.
Through: Game Radar
Author’s Note: I love unexpected results when training artificial intelligence. This experiment was undoubtedly quite entertaining.