[AI Challenge] Procedural Content Generated Challenge Disscusion...

It's live anyone who want's to join can. -> https://itch.io/jam/ai-pcg-level-challenge

Great Extra Credits intro to PCG

(an introduction to PCG in level design)

What about an AI challenge where we write an AI to build the best game levels.

So the PCG would generate the levels layout and place enemies, powerups ect.

So we could lock it down to a specific type and style of game for simplicity e.g.


2D platformer (making of Spelunky ->

)


3D shooter (Creation Process - https://80.lv/articles/the-creation-of-procedural-shooter-sublevel-zero/)

Or we could go for a more open challenge where people have a go at making any type of game they like procedurally and we play each others games to see how good the levels are.

Ideally games would be built to WebGL so anyone can try them and early entrants can get feedback on their games.

Although I would think that sharing approaches, resources and problems would allow people to learn about PCG with Unity.

And if you want to share your code you can but it is not needed in this form of challenge.

If we can agree a format we could use itch.io and launch it as a jam maybe even with an optional theme and use itch.io's rating system to rate our entries.

1. Open or Fixed Genre / Style?
2. Challenge Duration?
3. Optional Theme?
4. Any other items to consider?

Discussion...

 

> What about an AI challenge where we write an AI to build the best game levels.

This is a fairly big project that can easily take few weeks.

How much are you paying to participants?

 

No it's not that big a challenge I have written a simple PCG level generators for the Ludum Dare 48 hr game challenge...

And I think it would work best with a longer jam (weeks to a month maybe) so it would give people time to build, test and improve their PCG level generators.

And Unity has a PCG cave generation learning resource...

https://unity3d.com/learn/tutorials/projects/procedural-cave-generation-tutorial

And there are lots of PCG resources online.

No need to pay people who like to learn or enjoy a challenge as they don't need to share their code.

The challenge only needs a working WebGL build for people to try and rate how well their PCG generates levels.

Winners could put their PCG up on the asset store for other developers to make games with.

 

Dude, you're not taking procedural generation seriously.

Maze generators are not terribly fun, Interesting procedural stuff starts at planetary/city level and those take much longer. There are few things I'd tackle, but realistically speaking, that won't be any time soon.

Also, you marked it as AI. That's very different from normal procedural.

 

Tell that to the developers who made Spelunky or Dwarf Fortress!

 

Dwarf spent ten years in development, though, and handles much more than mazes. Not sure about spelunky.

 

Yah there's so many confusions about the use of terms. First off people start referring to every procedural game as a rogue-like or rogue-lite whenever something procedural shows up, and then people start referring to any procedural algorithm as 'AI'. Why can't people just refer to it as procedurally generated content? AI is mainly to do with a simulation of intelligence - you know, artificial intelligence - which can learn stuff dynamically and become able to gain new abilities etc on the fly. Generating procedural content can be decidedly very blind and procedural - like, just some ugly linear algorithm not even very object oriented yet alone AI. Anyway..

Getting this stuff 'right' can take a lot of time. It's easy to do something simple but you try to do a Spelunky or something where there's more design decisions and you're trying to create subtler gameplay or whatever and it's not just all 'random numbers everywhere'... that's harder to get right and can take a lot of time.

 

So writing a PCG algorithm that does good levels is not 'AI', even though you go on to say that coping with edge cases or scanning for dead ends is complex.

If you define AI as an automated system that does work that would otherwise take a person time (which is what most good AI does) then I think it is a kind of low level AI system.

Note: that you can get around some of the complexity of PCG by using modular blocks. Where the edge cases are accounted for e.g. doors with keys in a modular section.


A classic board game example of modular level design.

 

Very bad and cringe worthy, stop using random as a defning features, for FRAK sake!

Excellent overview, recommend.


I'm with Arrowx about PCG and AI

I have a "dump" threads about everything pcg, that's all the article I read over time, I read them all, even the redundant, small, ill defined, etc ...
https://forums.tigsource.com/index.php?topic=48415.new#new
I'm fond of hierarchical pcg system, where high level generator define/generate intents first and those are use to drive lower pcg system, basically like a language generator. I also try to make everything infinite and local as a way to generalize the rule set and make it efficient.

Pcg is hard if you try to tackle it as a single system with one level of "reasoning", nesting stuff give you more expressivity, most best pcg have that, like spelunky explicitly (draws path first, place room second, enemy last), or implicitly like pcg that use hand made model (the pcg create the level lay out, the hand mode bring low level placement). Even rogue had a implicit hi level intents, like each level grow in complexity and harder monster appear lower in depth, that is a progression intents, it gave structure to the middle randomness of each level.

Imagine if you had a pcg to define that hi level intent first (ie generating different type of progression) and a low level pcg (assemble monster stats and reward relative to difficulty, biomes, progression target, etc ...) it would be simpler to manage that the spagheti pcg we currently have. Ie design a pcg that thinks like a designer.

 

What if we don't define AI? This leads to dictionary wars and semantic arguments that produce nothing useful. ("System that performs work that would take time?" Well, your GPU is now AI, because drawing triangles by hand is time consuming. And this is why dictionary arguments are a waste of time).

If you drop the "AI" from the title (because it'll be possible to spend eternity arguign about what the hell it is supposed to mean) and concentrate on procedural generation, that could end up as a decent "challenge".

The reason why I don't see connection between AI and level generation is because AI solves problems continuously, while procedural generation usually produces whole level at once.

However, if a next room does not exist till player opens next door, and the type of room is determined by a script that attempts to achieve some sort of goal (like trying to murder player in most ironic way possible), then we have AI. It is similar to what L4D did with their Director script. However, even in this case AI is very decoupled from procedural generation, and concentrates on different set of algorithms.

Basically, two fields do not overlap much, and trying to jumble them together most likely will result in mess.

 

That's super arbitrary and I know many pro who will vehemently disagree, if you don't want a semantic war don't then push your interpretation lol

 

This could be a decent thread aside from the misleading title.

The current title is "[AI Challenge] Procedural Content Generated Challenge Disscusion"

It should be changed to "[Challenge] Procedural Content Generation Challenge Disscusion"

Note Generation instead of Generated and dropping AI from the title.

Procedural Content Generation can use AI techniques, but it does not have to. By including AI in the title, it implies to some people that this thread is limited to the subset of procedural content generation that does use AI, which is apparently not the case. So in an effort for maximum clarity, I strongly suggest changing the title.

 

I think your getting too hung up on AI...

Problem Domain: Go, Chess - Big news in advances in AI as computers out depth search the best human players!

Most AI problems are searching for a path or strategy in a game or they are looking for patterns in an image or video feed e.g. face detection.

Early AI research was just trying to find a better algorithm and build a faster/bigger computers to run it on.

It's only recently that AI research has started going into pattern recognition and machine learning algorithms.

Problem Domain: Level Design Can You Write Some Code That Will Generate Good Levels?

Is it AI, I think it is, you are automation a way to solve a problem. That's what most AI systems do solve one problem well, can the Go or Chess champion programs talk to you about your favourite stuff no, but maybe Siri can tell you where the nearest store for them is...

-----------------------------------------

It looks like the Open challenge could be the winner?

What about AI as an optional theme, as everyone has their own opinion of what AI is or does?

 

It's true that most AI are "basic" search or pattern matching, in different flavor. PCG is either "search" (ie test validity of structure like dungen, simulation) or pattern "building" (template and noise).

 

I don't know of anyone who uses ai to describe procedural generation. Ai (when used in conjunction with video games) is about in-game behaviors. The broader topic of artificial intelligence is something different. With tooling and systems, it's just logic. Very advanced stuff may use ml, but that is specific, not just complex. Procedural generation is just programming/ logic when you are talking about level / content builders.

 

Arrowx just posted a video that did just that about AI and pcg.

 

What you have defined is called automation. Please try and cite a definition of artificial intelligence that boils down only to "a thing that does work".


Now, as for the problems with a contest: simple generators alone can take a weekend. Just getting to the point where ANYTHING can be on screen is an entire weekend, and that's saying absolutely nothing about it's quality. More than anything though, the problems with PCG are all high level. I've seen nothing yet that actually shows any understanding of how to make interesting gameplay procedurally.

 

There is a course on AI at nucl.ai, done by veteran, you will be surprise about how AI is define, it's almost a joke among researcher because once a technique is sufficiently understood, it cease to be AI. People don't remind when database where super hi level AI, no it's just sql rofl.

 

There you have the Challenge part right there, how do you get randomness to coalesce into interesting level design?

It does open the further question of what makes good level design, and that depends on the game genre and control mechanics.

 

That's the wrong question right there, the role of randomness is to bring variation to an established order in order to mimick creativity. It's a standin for the fact we are all different and make different decision, randomness is just a way to simulate the process of making different decision. It's only as good as the modeling of the decision process, hence why spelunky is so good, it's made along the way we design level first, by starting with an overall experience (the critical path) and building around. Since the core of a level (here the progression as critical path) is solid, variation can only be solid first.

PCG should be the implementation of the design rules of a domain expert. What you are simulating is the space of knowledge of that domain expert, hence why ai and pcg are close.

That extra credits is damaging beyond relief

 

Randomness does not make level interesting. You need order. Randomness is chaos. Even if your level depicts chaos, there needs to be order in it, otherwise it will not be interesting.

 

10 Principles for good level design from a AAA level designer -> http://www.gamasutra.com/blogs/DanT...en_Principles_of_Good_Level_Design_Part_1.php

TL;DR Summary

Good level design...

1. is fun to navigate
2. does not rely on words to tell the story
3. tells the player what to do, but not how to do it
4. constantly teaches the player something new
5. is surprising
6. empowers the player
7. allows the player to control the difficulty
8. is efficient
9. creates emotion
10. is driven by your game’s mechanics

I think PCG should provide no 5. Surprising, not sure how well we can get it to provide the others.

 

Not exactly. He mentions procedural AI (which is a thing), and he talks about academic research related to "artificial intelligence" which is analytics and theory. There is a distinction. Practical application of pcg is not "artificial intelligence" and most definitely not "AI". PCG as term/description is complete. Adding AI or artificial intelligence adds nothing and confuses the term. And in the context of the proposed jam/challenge, ML is way out of scope. It would be like having "AAA MMORPG" jam. A maze/level/content procedural generation jam is practical and achievable. An advanced academic research based jam isn't. (at least not in this setting).

 

Scope. Scope. Scope.

Conceptually, a PCG jam/challenge is solid idea and could be fun. It is practical and has application. But if you over-scope it, it just devolves into a mess (as it has done before). Keep it simple and applicable and achievable. A good jam/challenge has constraints and boundaries to work within. (otherwise it's not a challenge).

Pick a theme/area : level generation(nav), or monster creator or visual decoration, something specific.
Set some achievable/measurable goals : like scaleable number of rooms, or maze that has no dead ends, things like that.
Define some structure. Like everyone uses the default 3rd person controller, or a base set of assets. Things that can easily be pre-defined to keep the focus the core challenge.

BONUS : Break it up. People are busy and have their own projects, making a huge open ended challenge is destined to fail. Why not make it series of manageable challenges. Challenge #1 : PGC environment (dungeon, city, rooms, etc). Challenge #2 : Build on the first challenge, PCG monsters/enemies. Challenge #3 : Add PGC Quests to what you have built. ... or whatever. Bite size, focused chunks that build on each other.

I would like to see this happen, and be fun and successful, that people want to participate in. (and not go they way they usually go).

 

Gah we are going in semantic war I pass, but ML is only a part of AI as a field, not the whole AI, you don't have to apply ML to get AI, although ML is where the current buzz of AI is.

I do agree with the next intervention about the scope of the challenge though.

 

The list of 10 are more things to consider when iterating on your and grading other PCG levels. Not things I would expect a developer to generate in a simple PCG challenge.

 

That list is too vague anyway, they are more motivational than practical, it say absolutely nothing about how to do a level design, just advise about the mindset to be in. The worse is thing like "create emotion", that has no value.

 

Knowing and using the correct definition for specific terms is important when communicating. You can't just call every computer program AI, because that is not constructive. Communities need clearly defined terms in order to communicate effectively.

 

You do realise that by saying that level design does not require intelligence you are being derogatory to level designers.

 

Long time ago I was messing with a 1st-person roguelike, where enemies were 3d letters. Looked like this:

Spoiler

Basically, without dealing with shadowcasting and invisible surface removal in C++, this stuff is basically a dungeon generator and is simple to implement, so there isn't much to learn about it.

Things I'd personally love to mess with would be city generation and planet generation, but that's not kind of thing I would be able to spend time on right now. Basically, I think that earlier "proof of concept double precision spaceflight" demo I made could be extended to planetary landing.

Also, like I mentioned there's no "AI" in it anywhere.

Monster generation ... someone could attempt building dwarf fortress "night creatures" in 3d, I suppose.

 

Red Corvette for Arowx

"Hey everybody, there is a red Corvette in the parking lot. Check it out!" shouts Arowx as he enters the office. Suddenly, everybody stops working and follows Arowx outside to check out the red Corvette.

Once everybody gets outside, Sally comments "Arowx, that is a blue Ford Mustang, not a red Corvette".

All of the coworkers look at the blue Mustang and agree that it is definitely not a red Corvette.

Unphased, Arowx starts touting all of the performance specs of a Corvette and shows everybody an article about Corvettes on his phone.

Then Arowx exclaims that Ford Mustangs are similar to Cameros and that Cameros are made by the same company that makes Corvettes, so (according to Arowx) Ford Mustangs are basically the same as Corvettes. Arowx also explains that they are both cars and have four wheels.

An engineer named Larry tries to explains the technical differences to Arowx between the blue Ford Mustang and a red Corvette, but Arowx quickly interrupts Larry and points to a car driving by in the street. "See?" says Arowx, "Another red Corvette. They are everywhere." as Arowx points to the green Toyota Prius in the street.

Realizing Arowx is Arowx, Larry turns and heads back into the building. As Larry opens the door to the office, he says "If every car is a red Corvette, then you might as well call every computer program AI."

Suddenly Arowx realizes that every computer program is AI and that everybody who disagrees with Arowx must be a naysayer. Then Arowx ran back into the office and started another thread on the Unity forums, this time with the title "Is Unity doing enough to stop the AI in red Corvettes from taking over the world?"

 

Absolutely nobody said that. What we said was "Arowx should not call every computer program AI".

 

"I pass, BUT" ?

Machine learning is normally associated with system that can be trained based on sample data.
AI could be thought of as expert systems.

Even though there's overlap, those are different things with different applications.

 

Now we need to take this story and turn it into central theme of AI+procedural "challenge" somehow.

 

Point out who said this.

I'll wait.

 

This thread isn't going well and I didn't help, sorry lol

 

And now you know why one should never start dictionary/semantic argument.

Anyway.

One thing I thought about was making a creature/design reach certain goal. For example:

The problem is, this is usually done via genetic algorithms and those are not AI.

 

https://en.m.wikipedia.org/wiki/Artificial_intelligence_(video_games)
https://en.m.wikipedia.org/wiki/Artificial_intelligence

Its not semantics, its context. An auto mechanic and a game developer both use the term "engine", though they are referring to two different things. As in the video, he qualifies the term with academic or research, except in one case where he doesn't and clear means behavior (game) AI. When discussing both the normal convention is "AI" for game behavior, "artificial intelligence" for the academic field. In either case, artificial intelligence (either form) isn't used as pcg methodology (at least not correctly). As @imaginaryhuman pointed out, the term pcg is self-defining. It would be similar to when people say "ATM Machine".

 

I'm interested in this, but I'm waiting for all of the vaguely annoyed responses to side issues to die down and some actual criteria for the challenge to be laid out.

Probably don't have enough experience to contribute meaningfully, but I want to see where it goes.

 

I stop taking bait, or else next time I'll find myself to google historical context and that won't go well either because even pro disagree lol. Bless Minsky.

 

Shout out for any resources and links to good PCG level generation tutorials, resources and examples.

Tom Bets - Sir, You Are Being Hunted (Made in Unity) small town generated using PCG.

No Man's Sky GDC Talk - How I learned to love procedural Art

Just some examples of developers and academics using PCG in games!

 

This has some fascinating information from an crazy expert in PCG.

 

What a confusion of terms. AI, unless otherwise clarified, refers to some sort of dynamic npc agent behaviour, and level generation is, well, level generation. Unless the idea is to develop a level generator that reacts to player behaviour then using the term 'AI' to describe a level generation algorithm is completely incorrect. Otherwise my endless runner tile-dropping algorithm would be 'AI', which it very definitely is not.

 

Please take all AI discussions over to here -> https://forum.unity3d.com/threads/discussion-what-is-ai.484270/

-------------------------------------------------------------

Good they should be gone now, where were we?

 

A link to a philosophical semantic argument is not the best way to clarify the terms of a contest. I still have no idea what you're proposing.

 

The contest is to make a game using PCG 'AI' system to build the level, then we can play and rate each others levels.

The discussion should be focused on

• Open or Limited genre/styles e.g. 2D/3D/platformers/fps/rts/adventure
• Jam duration, how long?
• Optional theme, should there be one and what should it be?

However the real goal is to learn and improve our PCG 'AI' approaches by sharing and providing feedback to each other.

 

Dude, turn off buzzfeed mode please.

I'd recommend to drop "AI" from the context, and concentrate on procedural generation only. This will be a decent contest/whatever. As long as you keep trying to stuff "AI" into discussion, it is incredibly unclear what the hell are you talking about.

Spoiler

 

Please take all AI discussions over to here -> https://forum.unity3d.com/threads/discussion-what-is-ai.484270/

Is PCG the indie game developer equivalent of bitcoin mining for nuggets of fun?

 

Wowsers, everyone in this thread needs to calm down!

Hey! My name's Mike and I organise the Procedural Generation Jam. We have some resources you might find handy, including a list of tutorials [http://www.procjam.com/resources/], a lot of videos from our live talk days [https://www.youtube.com/c/procjam] and some free art packs [http://procjam.com/art]. We're running again this November and we'll be releasing some tutorials this year too. We also have a zine that anyone can write for about stuff they're working on: http://www.procjam.com/seeds.

Ok sales pitch over! Let me know if you ever want to chat about any of the above (I don't lurk here much but I'm @mtrc on Twitter)

On the arguments here: PCG is considered part of AI by most researchers, and appears at every game AI conference I've ever attended. I did my PhD in it, in fact! I think it's really healthy to frame PCG as an AI problem here actually - it gets people thinking about systems that design things, instead of just a random algorithm here or there.

My recommendation would be to make a simple game, or get an open-source one, and let people write procedural generators for it. It might be less free creatively and people might find it constricting, but it'll be easier to compare and contrast the results. Some colleagues of mine actually tried it with a Mario clone as an AI competition: https://sites.google.com/site/platformersai/LevelGeneration

Whatever you do, let me know how it goes and what you get up to - I'd love to hear about the outcome!

EDIT - Also for people arguing over the definition of AI and what is too simple, you might enjoy this Wikipedia article: https://en.wikipedia.org/wiki/AI_effect

 

Where are:
1. Jam rules
2. Jam page
3. Jam date
?

Are there prizes?