[revelation]

we instead use features read from the game’s memory on each frame, consisting of each player’s position, velocity, and action state, along with several other values

So it's cheating, presumably knowing the opponents action before the animation even starts to play.

[scythe]

Smash is played on analog displays precisely so that the lag between RAM and the display can be as small as possible, usually 50 ms. In fact there's a 50 ms delay added to the AI for this reason. However, the AI takes no account of the fact that it takes about 230 ms for a signal to travel from a human's retina through the occipital lobe and motor cortex and activate the motor neurons in the hand. The AI can also generate input sequences that are nearly impossible for a human, such as the "dustless [i.e. perfect] dashdance".

But this is what a top player (who regularly beats both of the players tested in the study) looks like playing against a hand-coded bot:

https://www.youtube.com/watch?v=9qWHM8DNdr8

and this is what the humans eventually learned to do:

https://www.youtube.com/watch?v=be8UDlVuAl8

Even if you add reaction time, a big part of Smash skill for humans comprises accurately manipulating the analog stick. The computer can just declare any angle it wants; you're not having a fair competition until you build a robot thumb that manipulates a joystick the way humans do, IMO. Otherwise a character like Pikachu can recover perfectly every time.

[vladfi1]

The bots are given a very reduced action set. The Falcon master bot actually had only cardinal directions on the control stick - later bots were also given the 4 diagonals.

[vladfi1]

It does not see the opponent's actions before they take effect on screen, and the actual controller states are not part of the feature representation we used (though they actually are somewhere in the RAM).

[stagbeetle]

Part of the skill in competitive play is to be able to predict what move your opponent is going to do next.

Most mid-level players already have a good grasp of prediction, which is arguably along the sames lines of being able to know with certainty what action your opponent is taking a few frames before he does it.

Coupling that with pretty obscene frame-lag for Smash, it's not really that much of an advantage.

As well that competitive isn't really that impressive considering how limited your actions are by banning items and more dynamic stages (see: restricting RNG). In this way, it's nothing more than a simple chess-bot. Now, if it could actually take in complex environments and multiple tools, that'd be pretty next level.

[revelation]

You're wildly swinging between advantageous and not.

No, this is just playing games. The ground rules must be clear: you get the screenshots and keyboard input in every frame, as a normal player. If the resulting AI sucks, who cares? Failure is part of doing science.

[avbor]

I disagree with this. I don't think it makes sense to make the problem space harder and include translating screenshots into a resulting keyboard input when you can look at the memory and figure out the same information. That's an entire problem of its own, and I really don't understand why people keep ignoring that. Or is the issue that with the amount of information being supplied, it's too revealing? In which case, would limiting the amount of data being passed in to just, map, positions of the players, their current moves, and direction of the current move be okay in your book? Or is even that too much information to give.

[stagbeetle]

> You're wildly swinging between advantageous and not.

What?

> The ground rules must be clear: you get the screenshots and keyboard input in every frame, as a normal player.

Perhaps if you want to start from flawed assumptions/ want to create an AI that's tweak-able to appear as human. Which would be pretty useful and practical for other applications, but not competitive play.

We could go on and on about digital vs. analog, but digital is good enough for your argument and doesn't require you to spend enormous resources on a trivial pursuit.

This going in the direction of nonsensical handicaps. You don't give AlphaGo stamina parameters that artificially slow down processing speed. You give it all the tools it needs to beat a human player.

[erichocean]

> You give it all the tools it needs to beat a human player.

Okay, why not allow an NPC to just mess with the human player's actions then (blocking or delaying button clicks, for instance)? Surely, that falls into "all the tools", no?

IMO, the way you went about things isn't particularly compelling—your human opponents don't have white-box access to game internals, and if they did, guess what? They'd play better too.

So I agree with the GP: this is just playing games.

[stagbeetle]

> Okay, why not allow an NPC to just mess with the human player's actions then (blocking or delaying button clicks, for instance)? Surely, that falls into "all the tools", no?

Are you talking about physically delaying their inputs? As in from the controller to the main board? This would fall under the same category as a player hitting the controller out of his opponent's hand -- foul play.

> IMO, the way you went about things isn't particularly compelling—your human opponents don't have white-box access to game internals, and if they did, guess what? They'd play better too.

I'm not sure what exactly you're referring to here, but I'll respond to how I think you're trying to take this.

Source code wise: Yes. If the players had access to the source code the learning curve would be significantly shortened. Though, in due time, most would have figured out the mechanics fully, or within a short deviation, in closed source. A part of competitive play is this exact aspect. Players experimenting, sharing, and building up their understanding of the game. If the source was freely available to explore, most players would stick to the "show" part of the process, i.e working reflexes and learning combat -- what most elite players focus on (since they've mastered the science of the game already).