Kaladoga
Smash Rookie
Introduction:
Some of you may have heard of OpenAI, the artificial intelligence that played DotA and beat some high-ranking players. Well what if we got a neural net to learn Smash Ultimate? What could we do with that data? I have some ideas.
Ideas:
1.) Using a neural net AI to learn Smash Ultimate to the point that it could compete in a tournament and maybe even win.
2.) Using the AI to run many different battles against itself to collect data.
Parameters:
If we could somehow get our hands on an AI that could play smash against itself then perhaps we could use these parameters:
Match Rules:
For reference, a negative bias in this context means it's something that the AI actively avoids doing. A positive bias is something it goes out and tries to do. For the relative weights I'll add plus signs and minus signs respectively.
Training:
Imagine if we had an AI that was better than a human at Smash. Imagine how we'd be able to analyze the moves, we could then rank the characters based on the results. If OpenAI is anything to go off on, it'd figure out the meta by itself.
Trouble is, I'm not well versed in AI programming so right now this is just a pipe dream. That said I have seen something by OpenAi called Gym ( https://gym.openai.com/docs/ ) that might prove useful.
Some of you may have heard of OpenAI, the artificial intelligence that played DotA and beat some high-ranking players. Well what if we got a neural net to learn Smash Ultimate? What could we do with that data? I have some ideas.
Ideas:
1.) Using a neural net AI to learn Smash Ultimate to the point that it could compete in a tournament and maybe even win.
2.) Using the AI to run many different battles against itself to collect data.
Parameters:
If we could somehow get our hands on an AI that could play smash against itself then perhaps we could use these parameters:
Match Rules:
- 3 stock
- 6 minutes
- No items
- Final Destination / Battlefield type maps
- No handicaps
- Smash bar on
- No stalling
For reference, a negative bias in this context means it's something that the AI actively avoids doing. A positive bias is something it goes out and tries to do. For the relative weights I'll add plus signs and minus signs respectively.
- Positive bias: Winning a match: +++
- Positive bias: Making other characters touch the blast lines of the map: ++
- Positive bias: Dealing damage to enemy players: +
- Negative bias: Losing a match: ---
- Negative bias: Touching the blast lines of the map: --
- Negative bias: Receiving damage: -
Training:
- Run the game itself at an accelerated rate to make training faster for us humans.
- Run at least 10^6 iterations for a character against each other character in the game. (Minimum of 76,000,000 matches in total)
Imagine if we had an AI that was better than a human at Smash. Imagine how we'd be able to analyze the moves, we could then rank the characters based on the results. If OpenAI is anything to go off on, it'd figure out the meta by itself.
Trouble is, I'm not well versed in AI programming so right now this is just a pipe dream. That said I have seen something by OpenAi called Gym ( https://gym.openai.com/docs/ ) that might prove useful.
Last edited: