- 1
- 6 417 870
Peter Whidden
Registrace 14. 09. 2013
Training AI to Play Pokemon with Reinforcement Learning
Code:
github.com/PWhiddy/PokemonRedExperiments
Discord:
discord.gg/RvadteZk4G
Collaborations, Sponsors:
See channel email
Buy me a tuna melt:
www.buymeacoffee.com/peterwhidden
Sections:
0:00 - Intro
1:20 - How it works
2:54 - Let the games begin
4:04 - Exploration, distraction
5:46 - Level reward
6:38 - Viridian Forest
8:06 - A new issue
8:44 - PC Trauma
10:10 - Healing
10:45 - Gym Battle
12:43 - Route 3
14:44 - Mt Moon
15:54 - Map Visualizations
18:53 - RNG manipulation
20:07 - First Outro
20:26 - Technical Intro, Challenges
21:44 - Simplify
22:43 - Efficient Iteration
23:56 - Environment, Reward function
26:26 - Metrics & Visualization
27:46 - Future Improvements
29:24 - Run it yourself
32:58 - Final Outro
github.com/PWhiddy/PokemonRedExperiments
Discord:
discord.gg/RvadteZk4G
Collaborations, Sponsors:
See channel email
Buy me a tuna melt:
www.buymeacoffee.com/peterwhidden
Sections:
0:00 - Intro
1:20 - How it works
2:54 - Let the games begin
4:04 - Exploration, distraction
5:46 - Level reward
6:38 - Viridian Forest
8:06 - A new issue
8:44 - PC Trauma
10:10 - Healing
10:45 - Gym Battle
12:43 - Route 3
14:44 - Mt Moon
15:54 - Map Visualizations
18:53 - RNG manipulation
20:07 - First Outro
20:26 - Technical Intro, Challenges
21:44 - Simplify
22:43 - Efficient Iteration
23:56 - Environment, Reward function
26:26 - Metrics & Visualization
27:46 - Future Improvements
29:24 - Run it yourself
32:58 - Final Outro
zhlédnutí: 6 418 641
It stopped to look at the scenery, exactly like humans do It was petty and refused to press A for its defeat message simply just to not be told it lost It rage quitted avoiding brock because it lost too many times Your AI is just a literal human todler This AI is genuinely adorable
14:46 The music sounds so especially majestic here. What is it? Is it original from the game?
i hope they approved your master's degree dissertation
"rather well-endowed server"
Any chance the ai will ever beat the game?
This is how DJ Khaled would play Pokémon
omg so good
This is amazing, cool experiment and very well made video!
0:00 too many Ashes
Impressive!
There are better ways to do this than Reinforcement learning. Ideally with a game as complex as pokemon, we would program the rule set into the model.
the entire point is for the AI to learn the ruleset by itself
I like how the towns look filled with ai, I like to think that all the AI are in the same world and the people are just freaking out as to why there are 1000 of the same kid walking into the walls of there house
I still come back to watch this every now and again dude. I love it so much.
I'm absolutely dying for an update, I keep checking this channel every few months to see if there is another video
though ai looks like its doing what humans would do, it doesn't mean it's happening for the same reasons. AI writers should not cross over into philosophy without training, especially if you don't understand the difference between how you and your own program function.
Man uploads one video
bow before roko's basilisk
This might be the most interesting, fascinating and satisfying video I've watched on CZcams so far. Hats off, I'm looking forward to future videos!
"learning to play" and "learning a series of button combinations that will enable you to reach the end via throwing shit at a wall 10,000,000 times till you find which way to throw it and cover the entire wall", are not the same thing. this isnt AI, its brute force at the lowest shit tier level.
How do you think AI learns? Trial and error. Educate yourself before you make yourself look dumb lol
this is like complaining about musicians trying a bunch of different notes and chords before deciding which fits best lol
I really hope there was a universe where it did the mew glitch
wow amazing way of programming
when you got to mt moon and gave up i looked at the video run-time, saw that it was only half way, and thought "oh okay so clearly he's gonna decide to do more at some point" and then no. you didn't. instead you did nothing. i was at 27 mins in when it hit me you were serious. like... if you had to guess, do you think maybe people clicked this video to see an ai play pokemon or to watch you give up and pivot to a tutorial on how you made the video. come the fuck on.
Love the video, man! I can't wait to see what you do next!
You need to involve twitch in an active learning version of this
When comes next Video ?
when i heard "reward" i thought of giving the computer a cookie when i did something right. like training a dog 😂
How much can you tweak and micromanage the incentives before it's no longer "artificial"?
In future runs,make it so the Ai will have a fat reward for making a super effective move or boosting its Pokémon’s stats or lowering opponents stats
Visualization and editing is top notch!
I come back to this video every month or so because of how well it is done, I would love to see more about this, even in other games like Gen 2 or 3! What differences would need to be made to make those work? Would any AI ever be able to find the Regis? How long would it take for them to get past the sudowoodo tree? I would love more of this, amazing work!!
hey! all good questions. folks have been working on some really neat stuff, on multiple gens! there will be more updates shared eventually :)
it would have been hilarious if you used something from the red army choir for that opening gag. XD
it would have been hilarious if you used something from the red army choir for that opening gag. XD
Was hoping it would beat elite four eventually - good video
It looks like rollecoaster tycoon when you open the park. Can you emulate the game boy connection to have the AIs battle themselves?
In theory the goal of Pokemon is to fill the Pokedex, so perhaps it should receive increasingly higher rewards the more different & increasingly rare types of Pokemon it has caught. More abstract & long term than the goals you have so far, but perhaps at least as an extension to the rewards to better reward actual progression.
"Cheap junk pokemon" How dare you, sir!
Fantastic video. Thank you for this! Great visuals and explanation!
Bro drops a banger of a video then vanished
I would 100% watch the video of the first run where they beat the game. If they continued this and it eventually happened.
is this essentially the NEAT algorithm?
Not CNEAT that's for sure. Parts seem to be at least inspired by NEAT though.
@@Dalroc yeah he explained it more toward the end, i forgot what he said it was called. im gonna check this out more next week, God willing.
Bye for now, indeed.
Within a couple hours we took it's ability to enjoy the simple things in life away 🤣
this is not AI
Refusing to press the button lol "The only winning move is not to play"
is it possible to use your code to watch theai play from start to finnish ? btw i subbed :)
The only sad part of this video is that it’s the only one on his channel 😢 looking forward to more!
This is amazing. Where did you go? Could you make one to play Dragon Warrior Monsters 1 and 2 on GBC?
Its exploring the map like my roomba explores my house
Now imagine it's a 3D world and the characters main incentive is not to die.The rewards for reproducing as well as food and shelter are tremendous... this looked to me like the stepping stones of evolution uncanny to our own species. I'm not saying we're in a simulation, but I am stoned.
This is beautiful