Tool Assisted Speedrunning

And a small discussion on machine learning

What is Speedrunning?

• The first game ever to utilize speedrunning tactics was Drag Race, an arcade game released in 1977
• The goal of the game was to complete a drag race as fast as possible and in-house records were often displayed
• This brought about a new type of competition for playing games. Not just completing them but completing them faster than others

General Categories of Speedrunning

• 100%
• Completion of the FULL GAME. Depending on the game this could be all quests and sidequests. Or it might mean every collectible. It depends but there has to be an agreed upon 100% completion for all entrants
• Any%
• Completion start to finish. That’s it. Don’t worry about sidequests, you should actually probably ignore them. Straight line start to finish.
• TAS
• Our focus today. Tool-Assisted Speedrunning can be added to either of these categories and it’s when a computer is used to play the game rather than a human

Ok but that last part isn’t exactly true and you know it…

• Yeah, ok it’s not the whole picture anyway.
• So yes, the computer is playing the game but as we all know, computers don’t do anything without humans
• And AI isn’t good enough really (yet) to just play the game on its own (we’ll get to that)
• So… is the computer really playing the game?

Kind of.

• So, TAS and human speedrunning are very closely connected, in fact many human speedrunners contribute directly to TAS bots
• In particular, humans are good at finding glitches and exploits in the game… what does that mean for speedrunning?
• In many games, players find certain controller inputs or areas of the games that have… unintended consequences
• Sometimes you may find a glitch that is detrimental… sometimes it skips a level?
• Some great examples that come to mind are Portal, Super Mario 64

From one of my favorite games, Celeste

• There are a few things they mention in the video of particular note:
• “Human players go for these occasionally but the TASBot does it every time”
• “Human players usually take a different pathing but the TASBot doesn’t have to”
• “It has to take time to collect these because the TASBot doesn’t spend as much time on the ground as most humans do”
• In general, the consensus on how TAS behaves is that it saves time on humans either because it can do things more consistently or more precise than humans

​

What are they talking about “pathing”?

• A “path” is just a way to get from start to finish in any part of the level
• The tricky bit is juggling the layout of the whole level vs room to room
• Whenever they say “humans go this way but TAS just doesn’t” it has to do with the practicality of pathing
• Because the TAS can always do the right thing, it is practical for it to path the absolute fastest way
• Humans might not do it that way sometimes because the overall flow of the level is worse for them if they did

Pathing is the hardest part of TAS because sometimes you need to path in a way that only makes sense for a computer.

​

So sometimes, we ask computers to do it themselves.

Computers don’t even know what a game is

• The problem with computers playing games is mostly that computers don’t know anything about games (we can say much the same about other things ChatGPT and other LLMs can speak on)
• So you can teach it to use controls and you can let it move around but it’s not going to play with the intuition of a human because it isn’t one
• This can have both benefit and detriment
• Benefit: doesn’t know the rules so it can break them in ways humans might not think of
• Detriment: doesn’t know the rules so it might not even try to play the game to progress
• So how do computers play these games better than humans, then?

Carrot on a Stick

• While computers may lack intuition, we are able to tell them when something is good and something else is bad
• In this CodeBullet video, he teaches an AI to play the game Jump King where the main objective is move upward
• The main way that this is achieved is that he gives it a set number of moves it can do and rewards the one that gets the highest up in the level
• This would all work perfectly if the levels were designed so that the only progression happens upward but sometimes you have to go down to go up (*insert lightning mcqueen joke*)
• The way we get around this, is you place a reward at the place in the level that you need the AI to go, that way it doesn’t think that the most optimal route upwards is just jumping in place 5 times

TAS, ML, and Scrabble?

• Anyone who has done enough programming knows that computers are dumb as rocks
• The thing to keep in mind when using any kind of machine learning is that a human needed to tell the computer whether or not what it did was right
• There was a man who won the French Scrabble Championship and didn’t even speak French
• Just because you can score points, doesn’t mean you necessarily speak the language

Videos to check out

• Celeste TAS
• Awesome panel, it’s is a mix of actual human speedrunner, two of the main TAS devs and a commentator so very diverse explanations
• Michael Reeves letting his goldfish trade stocks
• This is the video I stole the Machine Learning screen grab from, it’s really good
• CodeBullet Jump King
• CodeBullet videos are fun dives into ML and other facets of coding
• French Scrabble Champion
• I know I just used this to make a point but it’s pretty cool
• Portal- Devs React to Speedruns
• A classic example of humans finding game bugs for timeskips
• Ape Out- Devs React to Speedrun
• There’s a great quote starting around 2:50 where the main dev is saying that they won’t fix the exploits because it invalidates all of these speedruns which I think is very cool