Analysing the AI of Super Mario 64 | Hacker News Reader

Analysing the AI of Super Mario 64

gamedeveloper.com

76 pointsdanso83 comments

I long for an RPG where I can have an actual conversation with an NPC. One with persistent memory, you know so that they remember that you murdered everyone in their family and stole everything they own. And not greet you every time like they've never seen you before.

Brutal example, but you get my point. Something that is ever changing.

or a pokemon that you actually "train" like you would a real animal, following all the principles of animal reinforcement (which you sort of get for free with neural nets)

I think black and white had that.

did not know, nintendo should take inspiration here

Black & White, a video game from Peter Molyneux in the early 2000s, had this. Unfortunately, the sequel simplified it to a set of slider bars, the company went under, and now Molyneux is focusing on NFTs and other scams.

The Creatures trilogy also had some pretty advanced AI, based on neural nets. They're pretty small by today's standards, but the paper below does a fantastic job of explaining how they are trained while minimizing the computational time.

https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.18...

that is absolutely fascinating. curious to read the paper, and see what the major obstacles were, or whether some newer techniques (like federated learning) could make the process more computationally efficient

I remember training your creature quite well. Teaching it to play "catch" with a boulder led to some interesting results, including:

- You were supposed to pet a creature if he did something correctly. Unfortunately, that same behavior also encouraged him to eat whatever he was holding. So I pet him after he caught a boulder, which then made him eat the boulder.

now that is seriously amazing gameplay mechanics

Fun(?) fact: Demis Hassabis (CEO of DeepMind) was the lead AI programmer for Black & White.

More than Black & White and Fable, it was the Creatures series that was designed around this ide:

Gameplay focuses on raising alien creatures known as Norns, teaching them to survive, helping them explore their world, defending them against other species, and breeding them. Words can be taught to creatures by a learning computer (for verbs) or by repeating the name of the object while the creature is looking at it. After a creature understands language, the player can instruct their creature by typing in instructions, which the creature may choose to obey.

A complete life cycle is modelled for the creatures – childhood, adolescence, adulthood, and senescence, each with their own particular needs. The gameplay is designed to foster an emotional bond between the player and their creatures. Rather than taking a scripted approach, Creatures series games were driven by detailed biological and neurological simulation and their unexpected results.

https://en.wikipedia.org/wiki/Creatures_(video_game_series)

I bought it on [well-known website that sells old, DRM-free games) but it was a bit... too old? So I never quite played it. So I have no clue how well the whole idea worked out.

I think this is the case in Gothic IIRC. As soon as you misbehave on some NPC, he will stop to cooperate with you (no more quests, no more trade, no more friendly smalltalk).

Which character do you have to kill to make him cooperative again?

Half joke, since too many RPGs have killing (or something equivalent) as their primary interaction. Geralt of Rivia must have extinguished the wolf population of entire countries, a band of Pokémon trainers would devoid an entire forest of animals to level up their pikachus, etc.

This is what I loved about Deus Ex, actions had consequences. Too bad it didn't catch on with later games.

Deus Ex only had a few moments with slightly difference based on your actions. The first one is the boss' reaction to your first mission depending on how violent you are in your approach. You get a few lines of praise or a chewing out and the weapons master will issue you some different gear for the next mission. That's about it. Later on, there are smaller difference I think like whether to manage to get the code word to activate a kill switch, whether you save a hostage... but the consequences are fairly minimal and confined each time.

I'm more impressed with the handling of consequences in Mass Effect 1 to 3. If you carry a single save game through all three games, the third game can become drastically different. One member of your party won't survive ME1 (it's your choice who that is) and the other one will substitute for them in ME 2 and 3. You can lose an almost arbitrary set of party members at the end of ME2. ME3 has a set of alternate party members to draw on when the main ones are gone. The fates of multiple NPCs can carry through all three games and end in very different ways (e.g. Karin Chakwas, Conrad Werner, Gianna Parasini). The controversy of the ME3 ending aside, the last entry of the series changes remarkably depending on your earlier decisions. Getting all of this straight must have been an enormous effort.

Mass Effect was still a massive disappointment though. So many choices, but they all converge into the exact same endings instead of diverging into wildly different possibilites. It really only determines how optimized the outcome is, the result is still the same. The only way you can fail is if you literally choose to do so at the very last moment.

Let's be realistic here: there's only so much that you can do on any realistic budget. A more elaborate and variable ending than what the ME series got at the end of ME3 would be a really big undertaking. I believe that the rest of the game would have suffered greatly from reduced resources if ending hadn't been as railroaded as it was. Maybe it's an unpopular opinion, but I got many hours of enjoyable gameplay out the series, so I can forgive an unsatisfying ending.

Deus Ex has a few more examples, like the fates of Paul and Jaime, but each of these things is explicitly scripted based on event flags, which explains why there aren't a whole lot of examples.

Right, I forgot about those! Still, these moments are designed to mostly be self-contained and have little to no interplay with each other. This keeps a nice lid on the complexity of the narrative.

A game's narrative complexity can easily get out of control when impactful player choices are allowed to become combinatorial. Implementing that many different outcomes quickly becomes infeasible.

In the interactive fiction world, where a great deal of effort has been put into complex, responsive storylines, a common structure is the "branch-and-bottleneck" -- allowing a variety of paths through a particular area, but all of the paths bunch back together into a single path at the end, with possibly a few major flags enduring for the rest of the narrative. As this article mentions, enforcing some structure on the narrative like this not only makes implementation easier but also tends to make the story more coherent for the player:

https://heterogenoustasks.wordpress.com/2015/01/26/standard-...

Fable does this pretty well.

Fable was the first thing that came to my mind as well. Not for how well it was executed but for how ridiculously oversold this aspect of it was pre-release. I think it was Fable that really solidified for me to take everything Peter Molyneux says with a Dead-Sea-load of salt.

Oh I see they came out with a new one called "Fable".

I was just talking about the original / the series in general.

I haven't played any after 2.

I'm very much talking about the first one (or two? It’s a long time ago now), actually, wasn't aware of the newer ones either! Don't get me wrong, it was a great game but not even half of what was marketed.

i read an interview with a developer who worked with Peter Molyneux at bullfrog etc. and he said that Peter would go into interviews and when he liked the suggestion of the interviewer then he would say that that was exactly how it would work in the game! So after he came back from a press circus he would present the developers with complete different ideas to implement than what they were doing before.

I'm very happy to have read this because now, finally, I understand what Molyneux was all about. At the very least a Marketing genius (among other things).

I feel like most modern RPGs have a tracker for your relationship with NPCs?

I think they're more talking about how in Zelda you can break into everyone's house, steal their money and smash all their pots, and they never acknowledge how you've violated them.

Zelda isn't really a role playing game. You don't get to speak. Few, if any NPCs can be attacked. Pretty much everyone is friendly. It's not exactly trying to go for realism.

> Few, if any NPCs can be attacked.

Sakon was a memorable case for me. I’m pretty sure he’s the only regular human you can kill in the whole series

> In Link's Awakening, the exorbitant pricing of the Bow at 980 Rupees may entice Link to steal the item from the Town Tool Shop in Mabe Village. It is possible to shoplift any item in the store by simply picking it up and walking out the door while the owner is looking in a different direction. However, stealing an item from the shop causes Link's name to be replaced with the title "THIEF." He will be called as such by every character in the game—even the Wind Fish—for the remainder of his journey.

Most people don't try it, but as a kid I felt a little slighted by the end game god beast refering to me by my one time misdeed :(

https://zelda.fandom.com/wiki/Thievery

The closest I've got to that is Disco Elysium.

You might like the Fable games.

I've been reading recently about the Language of Thought Hypothesis and the idea that conscious thought may be a production system. That is, it seems to work like a formal language or rule-based logic where a "sentence" (in the symbolic sense) may be extended or transformed by "rules" to produce new "productions" in theoretically infinite length and variety even if the actual performance thereof is bounded by limits of time and memory.

In contrast it seems almost all game NPC AI's are state-machine based. One could fit this into the framework described above by calling it a degenerate form of production system, where rules can be applied to transform a sentence (or symbol) representing one state into that representing another state, but none of the states are combinatoric and the rules just interact in a directed graph.

I'd like to see a game where, instead of just states in a state-machine, the NPC's could generate thoughts like a production system. Even if it wasn't connected up to complex action or a planner, it would be interesting if the NPC's could have inner lives based on a chain of thinking, like, "Mary loves Joe. Joe loves his dog. Joe's dog eats hot dogs. I wonder from where Joe buys the hot dogs for his dog." Etc. etc.

See: https://plato.stanford.edu/entries/language-thought/#ArguPro...

It's not an RPG and the interface is limited, but may I introduce you to Galatea?

https://ifdb.org/viewgame?id=urxrv27t7qtu52lb (you can play in your browser with the "Play On-line" button.)

Make sure to start with "help".

She gets annoyed at repetition. She'll react to things you do in the room. Responses differ based on prior conversation and some sort of internal mood implementation.

It's a breathtaking amount of work, but it's the very best NPC conversation I've seen.

There were games exploring in this direction in the 80s, but they were never sufficiently commercially successful, with the Ultima series being the most successful (and even with that, when Origin released Wing Commander they saw how much more profitable an actually popular game could be).

Perhaps give Trust and Betrayal: The Legacy of Siboot a try. Arguably NLP has reached the point where that game could be remade without the reverse-parser?

Which Ultima games are you thinking about? The series definitely was pioneering in simulation elements (day/night cycles and NPC routines in V, movable objects in VI, stacking and layering in VII), but the dialog system was always pretty simple keyword-based stuff, and NPCs mostly do not have any kind of memory. Are you thinking about how companions in VII will abandon or attack you if you do too many crimes with them around?

You are right that the dialog was simple and the NPCs basically got mind wiped every time you left a region. However, going back at least as far as Ultima IV, for the time within a region if you stole or killed, that actually had some repercussions. AFAICT the lack of persistence was due to technical limitations rather than a design choice.

IV is an interesting case because the entire game was built around repercussions for your specific actions, although the only long-term repercussion of doing comically evil things was that you had to spend a bit more time throwing money at beggars and overpaying at the reagent store.

V switched to a simpler karma system that mostly affected how badly you were penalized by dying and whether you were able to complete the game.

Reputation systems were pretty common in CRPGs starting in the '90s. In practice, though, a lot of those systems just boiled down to "does everyone in town want to kill you the moment they see you or not".

I think my original point that I made poorly was that in, say 1987 one would see that games were exploring the possibilities of the sorts of thing that the original commenter was hoping to see, but were held back by limitations of the hardware.

What we ended up with was 2-9 factions with up to 3 levels of response (hostile, neutral, friendly) which would be a disappoint to an optimistic CRPG player from the late 80s.

I think you could make really cool games around this too. I think the two main challenges and the reasons that most games don't work this way are:

1. It's often likely not fun. A big part of the joy of games is feeling empowered and free from all of the consequences of our actions in real life. Exerting power is often selfish and if NPCs remembered every single selfish thing you did, it would make playing the game feel as confined as real life. So much of the escapism of games is that freedom from consequences.

2. It's a testing nightmare. Games are essentially a ball of mutable state wrapped in a rendering engine. As any software engineer knows, mutable state is the largest source of bugs. Whenever you increase the amount of state the player can modify, your testing burden goes up significantly, potentially combinatorially. If NPCs remember your interactions, it means they are all carrying some mutable state around that can affect gameplay.

You've opened the door to an absolute testing nightmare where QA will file bugs like, "If you forget to ask Jane about her dog before talking to Fred about Jane, then four levels later you can't get her to tell you about her dog door and it's impossible to sneak in and find the key to finish the game."

It’s called real life.

To me AI is what is holding back video games the most. The graphics have advanced incredibly in the last 20 years, but the AI seems to be increasing on a logarithmic scale.

I believe it is a game design decision. We've had human like bots/AI in FPS games for over 20 years. It might be that most players enjoy the predictable patterns a simple game AI provides.

Right. A game has to be fun above all. Players want enough AI to keep things immersive but not so smart that it breaks the flow by being too hard to pass.

But people still enjoy playing multiplayer for the added challenge.

Bad AIs can really break immersion, for example in one of the Settlers of Catan games they'll refuse to trade with non-AI players. In civ5 harder AIs just get random levels of free stuff but otherwise are still rather dumb.

In FPS, most of the people playing are also bad, but their behaviors can be coherent. The enjoyable part is the random nature of your opponents and the context of the conflicts.

I suspect there's a sort of uncanny valley, where beyond a certain point bots become too complex to predict but too stupid to be interesting.

With deep learning maybe we have a hope to reach the other side where AI can add value.

AI could use a lot of immersive improvements, there is a lot of low hanging fruit above "hard mode enemies have 4x HP" or "the easy AI is an aimbot with a slightly longer reaction delay".

Nobody has invented effective programmatic mechanisms/patterns for general decision-making anywhere between if-else state machines and neural nets. The former becomes unmanageable spaghetti and the latter gives no theatrical guarantees.

So we're stuck in a kind of AI limbo.

Efforts exist in beating humans in specific games like chess; these games are often symmetric, i.e. all players compete on fair ground, and the AIs aim to replace human players. But most games don't want this! Most games seek theatrical asymmetry in computer-controlled agents. Most games want the illusion of cleverness without actually being clever, because most players don't want their gaming experience interrupted by undefeatable killer robots. Most games are power fantasies.

I posit that what we actually want is simply some programming pattern for managing complexity in autonomous theatrical agents. Some way to organize the creative will of the game designer easily enough that it democratizes independent developers to craft interesting behaviors.

I've got my own ideas but they're as of yet unproven.

Unity is pushing for Reinforcement Learning, with the necessary mechanisms for training, built into the engine.

That seems like a reasonable first step imho.

Fuzzy Logic as used my Massive might be what you are looking for

it would certainly make MMOs a little more interesting if raids actually required outthinking an intelligent machine as oppose to exploiting it's preprogrammed responses via rote trial and error

Not raids, but World of Warcraft's "Isand Expeditions" feature, introduced around 2018, does this.

They use an AI trained on PvP player behaviour, so the NPCs jump around, respond to teammates being injured, try to catch you off-rotation, etc.

In my opinion, it wouldn't work for bosses though, as part of the fun is exactly about understanding the tells and patterns. If you want to fight against intelligent enemies, just play PvP! I think if you play without addons like BigWigs/DeadlyBossMods you'll have much more fun, if what you're looking for is unpredictability.

Plus, would there really be that much of a difference between a machine-learned approach and a decision tree when every action the boss can perform needs to be pre-programmed anyway, and have predefined ranges/optimal uses? Maybe a different game could combine that with something like Magicka, where there are spell "recipes" which can be combined, so the AI would have more options to play with.

Another example is things like Dark Souls, the bosses are so engaging to its players _because_ you can learn to understand how they respond to your positioning. If they were a moving target, there would be no point in doing anything but learning the animations for each attack so you can dodge, so in the end they would be _less_ intellectually engaging.

that's a good argument. part of the fun is learning the AI behaviors. and in modern games it honestly would take a while for that to get boring, the AIs certainly are getting more complex

Actually you could have more advanced boss AI while still keeping the player ability to learn its "tells" if the animation system had access to what the AI is thinking. Lights that turn from blue to angry red when the AI is contemplating firing, giving the attack a mandatory charge-up time, etc.

The AI could be unaware of its tells or that it's telegraphing moves. Even more wild would be if the AI could be aware, so that maybe it tries to cover the status light or hide the weapon from view while it charges up. That in itself would be a tell.

The games industry has a lot of things in between state machines and neural nets. Since 2008, the craze has been goal-oriented AI planning (GOAP), which use a graph of states, and A* to pathfind towards a goal. And there's plenty of little tricks on top of this you can find in GDC talks. It's not without its tradeoffs though.

https://www.aiandgames.com/2020/05/06/ai-101-goap-fear/

> As detailed in Orkin’s 2006 publication at the Game Developers Conference whilst studying his PhD at MIT Media Lab, the GOAP system is driven by a Finite State Machine.

Emphasis mine.

Did you look at the FSM? It's an FSM, but it's not an FSM like traditional game AI; it only has 3 states! And in modern implementations, it's only got two. It's practically just boolean flag.

I feel like it must be intentional. The original Far Cry AI is far better than anything coming out these days.

Getting better? The original Half Life had better AI than HL 2. Should be more examples.

Did it? Or did it have better scripting?

HL1 had AI waypoints manually made by the level designer. If I remember correctly the human grunts were smarter, like covering eachother and stuff.

(I am not 100% sure though. Was my initial thought when playing HL2)

HL2 enemies especially the Combine Soldiers would zerg at you in single file. The marines in HL1 were way better than that. The ninja people were better still..

Ye good someone agrees I started to falter. From my limited HL1 modding experience, good waypoints were key for grunts. It probably made the AI way simpler to program with better results -- maybe an unsexy expert system -- which ended up better than the fancy general approuch. Wild guess

Enemy forces in The Division 2 were pretty clever. They cover each other, attempt to flank you, etc.

https://www.gamedeveloper.com/disciplines/designing-the-enem...

There's definitely a balance between "real" and "fun".

could it also in part be to the fact that graphics can be parallelized on GPUs, and AI behavior is inherently serial?

(and beyond perceptual tasks like vision, does modern AI benefit from GPUs that much?)

That's not what you want at all. An AI can easily have perfect aim in a shooter game. Designers need to handicap them to be fun. The same goes for many dimensions of games. You could make AI more competent but its usually less fun.

Even strategy games designed content usually come from you beating an opponent with a superior starting position.

What you really want is more immersive AI. Not in terms of the intelligence, but in the ability to do actions that interact with the environment. But this usually requires one off animation and gameplay coding so its not something that can scale easily. Things like hiding under something form a large enemy and thinking "he can't get me" but then, oh shit, turns out he is capable of crawling.

I think what he meant was a realistic AI, a perfect game logic for the role it needs to play, not a perfect solution for the task it does. Example: I want to have an AI for a monkey that behaves like a monkey would behave. And if that monkey sits in a drivers seat, I want the car to move like a monkey would drive, not like a perfect driver of cars.

But the issue isn't the AI, it's the interface that the AI has to the game. Monkey drivers is a little absurd so let's stick to monkeys in the forest. A real monkey would probably climb the trees. But if the game doesn't have climbing mechanics, then its unlikely that the game devs are going to implement them just so that the monkeys can do this. Making a behavior tree to utilize available actions is pretty easy to do. It's the implementation of the animations, game logic, and environment to allow such things to happen.

Little things like letting an actor turn its head to look at things will make the AI feel more alive than a more advanced algorithm for behavior. And you definitely won't see any particular gains by trying to solve it with ML brains.

>> An AI can easily have perfect aim in a shooter game.

I wouldn't mind that. I would be fine with an opponent like Scar in BSG:

https://en.wikipedia.org/wiki/Scar_(Battlestar_Galactica)

Something that's actually smart and really, really not just there for you to kill, for once.

Really? You'd like to play against a computer opponent that had instant perfect aim in a shooter? If you really want that go play some crappy online shooter full of cheaters with aim bots. It's not fun.

No, I said I'd welcome a game with an AI that's smart and not just there for you to kill. The bit about "perfect aim in a shooter" is your creative addition and I have nothing to do with it.

You quoted "An AI can easily have perfect aim in a shooter game.", and then said "I wouldn't mind that".

You may not have meant that, but it is what you said :).

No, that's what you want me to have said. I said I "don't mind" and you then went off on a tangent about instant "perfect aim" and I should go play a crappy online shooter with bots etc. And oh no, it's not fun!

You just decided you are going to crap all over my comment and crap all over what I said I'd like to see in a game because it does not agree with some kind of idea you have about "fun" in games.

More fool me for commenting in a thread about games.

The only conclusion I can draw here is that you've misread. I can only draw this because a) you're disputing a direct quote which is clearly visibile and b) you seem to think I'm the original person who responded to you.

I could be wrong here, but it may be worth going back and reading what was said again and the obvious inference multiple people took from your original comment, even if it wasn't what you meant.

Good luck!

My apologies for mistaking you for the original poster.

That is the exact part of my statement that you quoted?

I enjoyed the AI in Perfect Dark (N64), you could create a battle with a wide assortment of personalities (bot-alities?) like vengeance-bot (always out for revenge) or turtle-bot (try to get armor and hiding).

you should play the original FEAR if you haven't already. It has the best FPS bot AI that I can recall seeing. Part of it is that the AI often shouts out what they are doing so you get a better idea of their decision making process. It is a really great game.

Alien Isolation had some very in-depth AI governing the behaviour of the protagonist, and it really showed in gameplay.

Getting the feeling you're being actively stalked by a sentient creature is legitimately terrifying, and it's the first time I've felt scared by the intelligence of a game's AI.

https://www.youtube.com/watch?v=P7d5lF6U0eQ goes into some detail on the decision paths governing the AI.

>Getting the feeling you're being actively stalked by a sentient creature is legitimately terrifying, and it's the first time I've felt scared by the intelligence of a game's AI.

I've heard that the easier difficulty is scarier, because the alien appears less often so when it does it's more terrifying.

I actually don't think videogame AI has gotten much better since Halo 1. Maybe that game was just incredibly ahead of the curve, or that's as good as we can do.

Stupid games make players feel smart. Smart games make players feel stupid.

Players like to feel smart.

Reply on news.ycombinator.com