You could use the 100$ free credit that Linode gives you when signing up via a sponsored link. I do that all the time. You don't even need a credit card, just an email address that was never on linked before. This way, you could let your AI train on some more powerful hardware. Love the video concept btw, am really excited for a part 2
@@thedevatlasyou could also use a browser extension that makes temporary emails so you can just click the button and go instead if having to make new ones
I've been thinking of trying a similar project for last few weeks, of writing a bot to play the original Zelda for NES, so I wanted to share my 2 cents. I think what you are trying to do is not possible. In my opinion, Zelda games are too complex for a single reward function. I was thinking of having 2 networks (tactical & strategic). Where the tactical network focus on immediate actions like killing enemies opening chests, etc. While the strategic network focus on long term goals, like progressing the story, solving dungeons puzzles, etc. The tactical network would run near real-time, and take input from the screen and high level directives strategic network (which runs less frequently), to produce the final output keys to press. Anyway, I am no AI expert so I can be totally wrong about this. I just wanted to share my thoughts. Best of luck.
This is a fair point, I also think that it is harder than I thought going into it. However I did like how my bot was able to learn small things like exploring. You have a good idea and I think it would be interesting, but I agree that Zelda is too complex in general for 1 reward function. :)
I think you need to give the ai some human intuition. Even toddlers would want to leave the house and have learned outside knowledge kn where the exit is. Maybe you can slowly manually add information as it runs by highlighting something and say "thats a wall," "thats an enemy," etc.? I think your challenge is had this been mario, you could just reduce it to hold right, press a, press b and right, instead of giving it every button input. So as a result, exploring may as well mean explore absolutely every possible redundant combination. I apologize if i came off as harsh. I wish you the best in this project.
@@thedevatlas it would be cool if you could either one improve the ai of octoling in splatoon 1 or have an ai try to play it of that's even possible haha
support from swig
Nerd
You could give it a bonus when I reacting with objects for the first time and also make the bonus for seeing a new screen bigger
You could use the 100$ free credit that Linode gives you when signing up via a sponsored link. I do that all the time. You don't even need a credit card, just an email address that was never on linked before. This way, you could let your AI train on some more powerful hardware.
Love the video concept btw, am really excited for a part 2
Oooo I didn't know that. I shall take a look
@@thedevatlasyou could also use a browser extension that makes temporary emails so you can just click the button and go instead if having to make new ones
swig carries
Very cool 😎👍
Support from swig
underrated
I see you
I've been thinking of trying a similar project for last few weeks, of writing a bot to play the original Zelda for NES, so I wanted to share my 2 cents. I think what you are trying to do is not possible. In my opinion, Zelda games are too complex for a single reward function. I was thinking of having 2 networks (tactical & strategic). Where the tactical network focus on immediate actions like killing enemies opening chests, etc. While the strategic network focus on long term goals, like progressing the story, solving dungeons puzzles, etc. The tactical network would run near real-time, and take input from the screen and high level directives strategic network (which runs less frequently), to produce the final output keys to press. Anyway, I am no AI expert so I can be totally wrong about this. I just wanted to share my thoughts. Best of luck.
This is a fair point, I also think that it is harder than I thought going into it. However I did like how my bot was able to learn small things like exploring. You have a good idea and I think it would be interesting, but I agree that Zelda is too complex in general for 1 reward function. :)
I think you need to give the ai some human intuition. Even toddlers would want to leave the house and have learned outside knowledge kn where the exit is.
Maybe you can slowly manually add information as it runs by highlighting something and say "thats a wall," "thats an enemy," etc.?
I think your challenge is had this been mario, you could just reduce it to hold right, press a, press b and right, instead of giving it every button input. So as a result, exploring may as well mean explore absolutely every possible redundant combination.
I apologize if i came off as harsh. I wish you the best in this project.
Very cool, you plan to make more with other games
Yes! I hope to make some cool stuff
@@thedevatlas it would be cool if you could either one improve the ai of octoling in splatoon 1 or have an ai try to play it of that's even possible haha
How do you have the time to heart every comment
"he who hearts comments just likes reading them" - Sun Tzu or something
First
support from swig