Hacker News Reader

While the AI might be cheating by taking salient features from RAM rather than from pixel values, this is still an incredible feat. Just a few years ago we did not have generic algorithms that could take even salient features and self-learn policies to near this level this quickly.

View on HN

Yup, it's definitely an advantage to get all the correct values from the game state. But not as much as you might think; the vision portion of a DQN or similar trains quite quickly.

Plus, our bot doesn't have any clue about projectiles. We don't know where they live in memory, so the network doesn't get to know about them at all.

Can I ask what the feature set looked like? I always kind of wanted to do this with the Skullgirls AI, but never had the time while we were developing it. As a developer, I obviously had full access to the game state, but I'm still not really sure what the best way to represent that state to a neural network is.

It was just basic stuff like player positions, velocities, and animation states.

Getting them from RAM instead of the screen doesn't give you an advantage on (for example) DI or ledge teching?

Reply on news.ycombinator.com