Should I Release the AI Player`?

[ribot]

Is there really an AI player?
yes, made by me 2 years ago

Is it good?
no

Is it real AI?
yes (no, only machine learning), neural network

Does it play with any intelligence/skill?
yes, but see quesiton 2

So why should I release it? Could be fun to see what kind of intelligence it develops and how it evolves. There could also be visuals and discussions of problems of making a good ai.

[A.K.B.]

ribot please respond to my rpg coment

also ask jon if AI-cycle is okeallowed

[kuchitsu]

Go for it man! And don't beg for 10 likes or whatever, that feels a bit attention whorish. Just do it!

Btw Stini (?) is also working on something similar, not sure if he released the code but I've seen some recs and they were really fascinating.

edit: oh, or are you asking because some people might be against the idea? in that case sorry for the whore remark.

[AndrY]

make a poll with some 5 suspected ppl and we will guess:)

[pawq]

I think this might be better discussed than voted, do you think there's any potential risk ribot? Could it be better to release some offline recs like Stini has done first, before letting it battle, if it's good enough?

One thing for sure: it should be clearly labelled as a bot/ai in its nick, so it's obvious to everyone.

Pretty exciting that there's another one developed in the community tho! Soon elma AI league :D

[ArZeNiK]

if it is irreproducible for online use making it seem like a normal player (no matter if superior or not), then yes please release it i am very much interested

[ribot]

AKB: I can't remember anything about RPG now, and ai player won't work on other elma version because it controls bike thru reverse engineering of eol.exe
Kuchi: the reason I ask for interest is because it's quite a lot of work to make this release, and without any attention why should i put in the efforts`?
AndrY: I would make poll in here it if were possible
pawq: there could potentially be risk if i release code, but I doubt my code is anything of use to someone with skills like that

Everyone:
Just so you know there is nothing fascinating about this. It's a very simple machine learning bot that becomes better by itself. But it would not beat the worst noob in battle, even after many hours of training that level. Or at least me with awful elma skills will beat the bot at first try.

To me it was still exciting to see how it would play in the beginning and how it would learn.

The bot doesn't use volt keys, as I gave up before it could learn to play with those. However, it does look like it has some real elma styles.

[jonsykkel]

release it, awsj needs someone to compete against in balles

[ribot]

When I wrote this topic I didn't know there was Stini's AI. I don't think there's any point releasing it.

[ribot]

Anyway, no one has protested that I release the code.

[mena]

I also made an AI player but she is stupid, can't finish int01. We could make an AI WR table maybe.

[ribot]

I think mine could finish warm up.

I guess the main difference from mine and Stini's would be that mine was playing elma in real time, so you can imagine what a waste of computer power I have done. And that his was actually impressive.

Now this topic can be delete.

[pawq]

That someone else made something more impressive doesn't mean that your stuff is worthless =) Would still be fun to see!

[ArZeNiK]

please dont compare yourself to other people, creating an ai that can play elma is already a feat by itself no matter how advanced it is.
go for release imo!

[ribot]

oke here is vid: https://youtu.be/wPPsbdjJC78

[ribot]

I made new AI that could finish 01-04. You can see quite clearly that it's a computer that is playing.

01: https://elma.online/r/grkqqpyjec
02: https://elma.online/r/7y7ei8mchw (better time only using gas)
02: https://elma.online/r/890bah8ldx
03: https://elma.online/r/goqsm9db2e
04: https://elma.online/r/gqhuj6dk77

and pipe: https://elma.online/r/eokowdyohc (episodes: 1730, time: 4.59 seconds, elmatime: 5972.76 seconds; inception speed 1299X)

[ArZeNiK]

looks decent, actually

[ROKKEBOL]

Very good recs imo! 1-3 playing like a real human nab player, but stronger, no panic volts osv.
4th rec is quite pro for a nab, resolute moves at killers and everywhere.
5th rec showing uphilling skills, also noticed wheelpoo and cool end.
Me - like the AI Player

[iCS]

Impressive and amusing replays. Good to see such complex moves and routes.

[kuchitsu]

This is super fascinating! The pipe rec is sick. I have a question: does your AI essentially learn each level "from scratch", or can it gain knowledge from one level and then use it in another level?

[ribot]

Thanks for your feedback!

Kukitsu: it learns from scratch

[Xiphias]

Wow that is insane Great work Ribot!
Will your AI player play online some day? Is that possible (if we ignore eol anti-cheat rules)?

[kuchitsu]

What happens if you give it an auto level? Can it eventually realize that the best tactic is to not press anything?

[ribot]

Wow that is insane Great work Ribot!
Will your AI player play online some day? Is that possible (if we ignore eol anti-cheat rules)?

It's possible. So far it didn't perform very well but it could be fun some way.

What happens if you give it an auto level? Can it eventually realize that the best tactic is to not press anything?

TL;DR
It would just play randomly getting closer and closer to the flower until it finds a way to finish. After that it will make some improvements but it will still suck as hell, unless the level is extremely simple. I don't think the best way to finish auto levels is to do nothing, but if that would truly be the best way to finish (meaning getting the best time) the AI could find that out depending on the level and scoring system. In my case I don't think it would do nothing becasue some random stuff would not cause a worse time so the AI would never realize it can do nothing.

The AI did beat me on one level but it did not beat the best players. On this battle https://elma.online/battles/152353 pressing only gas, the AI could beat me getting 7.62 in no time at all. I would say perhaps faster than Spef, but then it wouldn't improve to anything better than 7.60. It could get 7.44 pressing all keys. One way to estimate the simplicity of the level is that random play at this level pressing only gas resulted in 7.77 (where I thought i had implemented an AI that contained a fatal bug).

Explanation:
From the way I see it (I'm really a nab at machine learning). First of all the quality of performance is dependent on the accuracy of a score system. Every attempt to play a level (episode) will give a score, and the AI will try to get the highest score possible.

A simple scoring system is to measure the distance from the bike to the flower: The shorter the distance, the higher the score. And then a bonus for finishing the level, and a higher bonus for a better time. This is enough for teaching the AI to finish some levels and make some improvements.

But how far can it go? How well can it perform?
Well, easy levels like internals 1-4 have big scopes of allowing non-fatal mistakes. In other words you can finish those levels in endless ways, even if you make mistakes. So you don't need to find a difficult and accurate combination of moves.

Now the AI tries things randomly. So the probability of it making some insane and accurate combinations is perhaps feasible, but would the score system be accurate enough to tell the AI that it did good? For this to be true the scoring system would have to take into account that a bad ride with a very good part is better than a descent ride. Things like punishment for death and distance from flower are no longer enough. The distance from flower has to be upgraded to take into account all parts of the kuski that can touch the flower.

I simply could not come up with a good enough scoring system to make the AI get good rides.

[pawq]

Sick stuff! That last killer dodge in twin peaks is dope, especially with no brake :D And sik piping skiallz!

[Bludek]

I was wondering, if machine learning could be useful in SL.

Maybe human can drive the parts that are hard to learn for AI and AI can brute force some stuff like killer dodging in int39 Haircut wr-style or idk.

AI so far seems to have very few advantages over human (maybe turning in rapid succession and killer dodging?), but they might be not negligible.

Is it possible to even use AI in SL? Does AI really have any advantages over mans or am I just making stuff up?

[ribot]

First of all it depends if you mean my Ai or a state of the art Ai of 2020 (or say 2030). It's possible we will see the day when AI can beat all wrs.

I guess it could be used with saveload, but I'm not familiar with those file formats.

Maybe the current AI's do have an advantage in piping. In piping it's not so important what you did earlier, compared to a brutal volt where you have to do some different moves in combination for it to work. Maybe only straight pipes, but perhaps difficult ones. For example driving a stilt bike (modded) is super hard, yet AI made the pipe using it:
https://elma.online/r/drnj90fzlp

A combination with human players is possible and I think the most realistic approach until a lot of AI progress is made. An AI could brute force a small part of the level if there is some logical way to tell which ride is better than the other. Things like taking the apple or not dying might work, but as I said earlier I couldn't find a very good way to do this in general. Another approach is to create a training lev where the player will for sure die unless doing something really good.

I see possibilities but not that easy to implement.

[gimp]

The over and under rec does appear to have some inhuman element to it. Like it is doing tons of turns to try and maximize some formula, doing flips in odd spots also to land on wheels that i dont think a human would do intuitively. Was interesting, wondering if it has potential to beat humans.

[ofta]

nice ai, i got an idea is it poss to tell the ai to keep the head as close to the ground as possible while still collecting aples and flower? that way it mb stops doing those 360 volts and perform better?

[Xiphias]

I think the AI has great potential to be used in conjunction with auto generated levels to check if they are finishable.

[ribot]

The over and under rec does appear to have some inhuman element to it. Like it is doing tons of turns to try and maximize some formula, doing flips in odd spots also to land on wheels that i dont think a human would do intuitively. Was interesting, wondering if it has potential to beat humans.

All the recs are pretty much playing randomly until being able to finish. So those moves were better than the prior, according to its previous experience.

nice ai, i got an idea is it poss to tell the ai to keep the head as close to the ground as possible while still collecting aples and flower? that way it mb stops doing those 360 volts and perform better?

This line of thinking is how you progress. You get an idea, implement, and watch the outcome. However, imo, as close to ground as possible means volt to death very soon. There could be a reward for being closer to ground, but I don't think it makes sense generally. You still have to balance everything beneficial with the reward of coming closer to flower, and yet not punish enough to make the kuski just freeze in fear of doing something wrong.

I think the AI has great potential to be used in conjunction with auto generated levels to check if they are finishable.

Uhm well... if the level is super easy. It can test things out, and learn very fast when I modded the bike. There's quite a difference in seeing its skill from the default bike compared to a bike that you're completely noob with.

[Xiphias]

I think the AI has great potential to be used in conjunction with auto generated levels to check if they are finishable.

Uhm well... if the level is super easy. It can test things out, and learn very fast when I modded the bike. There's quite a difference in seeing its skill from the default bike compared to a bike that you're completely noob with.

Does your AI play in real time? If so, could you use a modded elma.exe where time runs faster? I remember the MotoCoin project where you had to play an elma-clone to finish levels to generate hashes for the coins. In that project someone made a bot that would finish almost instantly (but real world time of run was 5-20 seconds I think) since they could just increase the time-clock speed by 100x-1000x or more.

[ribot]

My fastest algorithm runs around 1500x faster than in realtime. Also my computers are like 10 years old.

[kuchitsu]

I guess ideally you would want to modify the exe so that it doesn't actually render anything, which should save quite a bit of processing power. Unless you're already doing that.