LogFAQs > #941668978

Sonicplys posted...

Why do extra work. Fuck Manual Transmission Cars. In video games and real life.

The idea is basically that an automatic cannot do it as well as a well-trained human. This is, naturally, pretty much just important for racing.

Computers can certainly calculate things nicely and ensure smooth transitions, yes, but they cannot match a human pretty much for the sole reason that they don't know what the human wants to do next. They aren't looking at the track. The human sees the next part, and is anticipating they way they will be increasing or decreasing their foot's push on the gas pedal. In different circumstances, they may start out doing the same thing with their foot, but whether or not they want to shift or not may depend on what they expect to happen next.

Automatic transmissions cannot know these things. Even once we have safe, self-driving cars, the AI for these cars will not yet be advanced enough to know these things well enough to handle the transmission in the way a human can.

The most impressive game-learning AI yet, AlphaStar (yes I say this is more impressive than AlphaGo), still was using a simulated setting and thus was able to greatly accelerate the game as it played against itself and its variants again and again. Whereas taking the Alpha series AI approach to racing would be much more difficult, with you needing a very accurate simulation of the driving physics and at least a reasonably realistic set of graphics to look at.

AlphaStar, btw, is an AI that plays Starcraft 2 multiplayer. It learns by playing against itself at a highly accelerated speed, and has here and there been released into the ladder to play a set number of games under human supervision (mainly to join the game, and to surrender when necessary because the AI doesn't actually know how to join or leave the match). It hasn't reached the point where it can consistently beat the highest end players, but it does well and sometimes comes up with some interesting strategies. This AI does not see the game from "inside," but rather sees the game the same way you would on a screen, and inputs actions to the game similarly. As such, it is effectively playing the game the same way a human would, looking at the screen and making decisions based on this.

As it has gotten better, the people involved with the project have been applying stricter actions-per-minute limits to it, making it play more and more like a human player. This does mean that some of the cool super-micro-managing that earlier builds pulled off (like that time it basically was ordering individual protoss Stalker units to blink away with high speed micro before they died) is not seen so much in the later builds, and its greatest strength has shifted into the macro game, efficient resource gathering and expansion, etc.

You can search youtube to see people commenting on replays of games AlphaStar played if you are into RTS games at all.
.

---

I am the Hunter of Topics. My post never fails to kill its prey.
*pounces* Nyaa!