Comment by seanwilson | Hacker News Reader

Comment by seanwilson | Hacker News Reader

I tried making a random text adventure generator a while ago, where the idea was I would fill it with items and item combinations, and ask it to look for the most crazy/interesting chain of item manipulation puzzles e.g. key opens chest, to get the matchbook, to light the dynamite, to blow open the safe, to get the money, to buy the ticket.

I thought the idea was really promising at first but my finding was there's just not that many interesting puzzles based around real-life item interactions, so it wasn't fun to play (without going further by adding a plot, NPCs and graphics).

If the item interaction is too everyday, it's too obvious and boring (like a key to open a door), and if it's too obscure it feels unfair (like a blowtorch to the neck of a wine bottle to open it), so it has to be somewhere in between and there's only maybe a few hundred types of interactions like this. When you look into it, you'll notice the same item interactions reused in lots of adventure games too e.g. bolt cutters + chain, torch + batteries, spade + ground.

Maybe sounds obvious when you think about it, but it wasn't obvious when I was prototyping.

I think this is one of the reasons escape room games devolve into obscure logic puzzles. And also why adventure games got criticised for having super obscure (moon logic) puzzles. And also why a lot of adventure games feature things like time travel and magic and NPCs, as it lets you introduce new rules for item interactions.

Another aspect is filling the game world with lots of items, locations and general red herrings so that the connection between items that is usually obvious isn't immediately obvious because you can't keep them all in your head.

The bugs were funny though. Like you had to tell it you can't put a car in a tree, that you shouldn't bury bread in the dirt, and a vending machine can't dispense vending machines.

> The bugs were funny though. Like you had to tell it you can't put a car in a tree, that you shouldn't bury bread in the dirt, and a vending machine can't dispense vending machines.

My god, leave those in!

It does sound fun, but there would have to be some layer to control the amount of wackiness, or it would feel random to the point of meaninglessness, like "procedural oatmeal"?

You may want to look into the nintendo zelda chemistry engine

Any good links? I briefly looked up that it models fire, electricity, cutting etc. interactions, but if you put this in the frame of a text adventure rather than a 3D action adventure, would it be fun? e.g. typing "shoot arrow at explosive barrel" isn't as engaging as aiming the arrow yourself.

Sibling comment mentioned Zelda but really this reminds me of the Zelda Randomizers.

Fixed set of items in a fixed set of locations. Certain locations need certain items to access. So therefore randomize which items are in which rooms while still being able to complete the game.

Then they have a door randomizer which changes where doors warp you to - effectively changing the map and potentially the order of items needed.

> Fixed set of items in a fixed set of locations. Certain locations need certain items to access. So therefore randomize which items are in which rooms while still being able to complete the game.

So this part of the puzzle generation isn't that bad, you also need to make sure you can't make it unsolvable by destroying/wasting items or making them inaccessible (or allow this and call it "Sierra mode"). Making the item usage interesting is the hard part, but that's not as much as an issue for a Zelda game.

I figure its mostly a graph problem. I do like door randomizers, esp ones that don't have bi-directionality.

Yeah, it's essentially https://grumpygamer.com/rtmi_pdc/ where the items are nodes and the edges are item interactions, and you're trying to carve out a subgraph for whatever your definition is of a good adventure game.

I think you've reinvented the game Scribblenauts.

Does it do interesting chains of puzzles anywhere (more like Zork)? From what I've seen, it looks more like isolated puzzles where you summon any object you can describe to solve the problem rather than having to make do with what you find.

I also experimented with randomized text adventure game development back in the day but went a different route. I created an n-gram markov model scraped off hundreds of old I.F. games. The maze would be generated, and the markov model would be used to generate the description of each of the rooms. I then used a keyword/POS extractors/etc. to extract a structured well-modeled data representation of the room in terms of items, monsters, etc.

It worked kind of well, but it was more of an experience in exploration (almost like a MUD) rather than a cohesive linear game.

Of course, much later down the line, games like AI Dungeon really dialed up this concept.

So I've tried playing with LLMs for this and the problem is you can talk your way around puzzles because you can't guarantee it'll stick to constraints. If you want something more like a traditional adventure game, it would probably work better to algorithmically generate a cohesive linear game, then use an LLM to fill in the gaps in the descriptions, story, NPC conversations and images.

I don't see how you could avoid the LLM saying stuff that implies workarounds to puzzles though e.g. you add a knife to the game that's meant to be the only object that can cut a rope, but the LLM describes a rock as sharp.

This is my experience as well. It's why LLM dungeon masters can tend to fall flat since they have a hard time distinguishing between a genuinely good alternative solution vs somebody clearly cheesing it.

If you want to force a solution, you have to add an additional prompt check - e.g. "Is the solution provided by the player properly making use of items only within their inventory (A, B, C -> from their actual inventory) and does the solution adhere to one of the following acceptable solutions (X, Y, Z)?"

This helps mitigate abuse but it can also unnecessarily constrain the player, so YMMV.

I just don't see how you can stop it from going off the rails though. For linear adventures, you have to carefully pick which items and actions are available to avoid breaking puzzles e.g. if the LLM lets you break a window which creates a shard of glass, you could use the glass to cut a rope, when you were suppose to use a knife. It would probably be able to come up with decent reasons to explain why you can't do stuff though haha, which is often a part of good puzzle design e.g. "I don't want to draw attention by breaking the window".

I imagine there's also a potential exploit if you're ever allowed to name your pet, and the LLM isn't shielded from ever knowing the name.

Puzzles are to adventure games like frosting to a cake.

It’s expected to be there as a convention, but you want to keep it to a minimum because too much spoils the whole experience.

The story is the point and the puzzles are just an accepted contrivance to make you stay longer within the world. They’re not objectively fun, like frosting isn’t objectively good.

I partly agree, but there are puzzle heavy games with throwaway plots and people do mention they like games because of the puzzles, so I thought my observation was non-obvious.

People still talk about the death of adventure games being caused by moon logic puzzles, but I don't hear people mention that part of the problem is the design space of item based puzzles is more limited than you'd think (I mean non-magic, non-NPC stuff) and has been mostly explored.

While the story is definitely very important, a point-and-click adventure game without puzzles is basically just a book where you turn pages by clicking the mouse.

Of course there are good puzzles that integrate well with the world (like reading from the world's most boring book to help the oyster fall asleep in King's Quest VI), and bad puzzles (cough arbitrary sliding tile puzzle cough).

Reply on news.ycombinator.com