Using Python to build an AI to play and win SNES StreetFighter II with machine learning
video description
Date: 2022-03-14
Related videos
Comments and reviews: 7
Howard
The one thing missing with these rewards training system is that the system can't generate good intermediate rewards to target. When humans train, we identify the intermediate steps needed to get good results like we practice a certain combo and then optimize for the opportunity to connect them in the main game. We get rewards just for timing those things right. But AIs don't specifically target for them. That makes it so it wanders around all the time. So even with a lot of training power, a lot of training power is wasted training for useless things. It's true that it helps finding out of the box solutions but most of the time you get unoptimized solution. You can see that with just that running 100m (forgot the game name) video. He used input from a master as a priori for the engine to optimize. This way the AI doesn't have to wander around and settle for a local minimum that's very far from the globally optimized solution.
reply
The one thing missing with these rewards training system is that the system can't generate good intermediate rewards to target. When humans train, we identify the intermediate steps needed to get good results like we practice a certain combo and then optimize for the opportunity to connect them in the main game. We get rewards just for timing those things right. But AIs don't specifically target for them. That makes it so it wanders around all the time. So even with a lot of training power, a lot of training power is wasted training for useless things. It's true that it helps finding out of the box solutions but most of the time you get unoptimized solution. You can see that with just that running 100m (forgot the game name) video. He used input from a master as a priori for the engine to optimize. This way the AI doesn't have to wander around and settle for a local minimum that's very far from the globally optimized solution.
reply
nesterzhzhot
If this is for beginners, than I don t like it, because there are a lot of info, that I needed to know, but they didn t tell.
If not, than why they gave so many facts, that pro python programmers already know?
Anyways, I liked this video, good job!
reply
If this is for beginners, than I don t like it, because there are a lot of info, that I needed to know, but they didn t tell.
If not, than why they gave so many facts, that pro python programmers already know?
Anyways, I liked this video, good job!
reply
70ME3E
whoa what, the presentation just ended when I thought it was just getting warmed up..
where was the main model's architecture shown? you talked about some boring infrastructure setup stuff and about your cute on site display but not the main part??
reply
whoa what, the presentation just ended when I thought it was just getting warmed up..
where was the main model's architecture shown? you talked about some boring infrastructure setup stuff and about your cute on site display but not the main part??
reply
nakobots
Would have liked to hear more about the observation space. From the talk we heard x position and health bars are observed. I wonder what else about the opponent's action is known to the agent.
reply
Would have liked to hear more about the observation space. From the talk we heard x position and health bars are observed. I wonder what else about the opponent's action is known to the agent.
reply
Steven
A 30-minute presentation without actually seeing the results of their efforts. I wish Google's AI was clever enough to not show me these videos.
reply
A 30-minute presentation without actually seeing the results of their efforts. I wish Google's AI was clever enough to not show me these videos.
reply
crc_error
Fascinating. Fighting games and artificial intelligence are both interests of mine and I'd love to see more development of AI in these games.
reply
Fascinating. Fighting games and artificial intelligence are both interests of mine and I'd love to see more development of AI in these games.
reply
MavVRX
What does the agent do when the player swaps sides? Was it trained on both sides or was the input and controls flipped to handle that?
reply
What does the agent do when the player swaps sides? Was it trained on both sides or was the input and controls flipped to handle that?
reply
Add a review, comment
Other channel videos
2:50:8
