Advances in Artificial Intelligence for next generation, & advanced AIs in videogames

[Shifty Geezer]

Ie. No really enjoyed playing Star Control 2 on the hardest difficulty. F the Sylandrome

This is a good case where human-like AI can be a disadvantage. I've improved the AI in my ionAXXIA. Initially I was looking at dumb AI like an arena shooter, but felt the original 1 on 1 premise of Star Control was necessary to require specific ship designs. As I was working on AI, and thinking through how a ship could determine when it's best to coast backwards and shoot the other way, I realised that I was pushing the AI towards stalemate. It'd end up a war of attrition with mistakes resulting in defeat, such as a rogue asteroid hit here or a slightly over-enthusiastic turn there.

So I'm limiting how far I'm going with AI for now. It needs to be varied enough for varied gameplay, but ultimately very defeatable so players can kill dozens of ships. Looking at solo games, it seems it's the sense of success and power that many gamers want. There's been times when coop gaming with friends I've pushed for a real challenge (CON level 1 chars at hardest difficulty) and they get bored and frustrated with the long combats and prefer easy button mashing.

There's obviously scope for smarter challenges as shown in the Souls games et al, but I reiterate that most game are going to be more fun for most gamers with rather dumb AI for the reasons Esrever states.

On the subject of how to get smarter intelligence though, the obvious route these days would be machine learning. Send huge amounts of gameplay data from real human players to The Cloud and have it churn out strategies that can be folded into a simulation. I've no idea how that would be implemented though!

 

[iroboto]

You guys are thinking about game AI as what makes the enemies behave strategically smart, but you can think about it in behaviors that make the game more engaging. Enemies that have different behavorial paterns, that are well comunicated to the player through animation, taunts, etc. Teams of enemies that dynamically react and adapt not necessarely to be more competent at killing the player, but to create more interestin gameplay scenarios.
Those are the challenges of AI for these types of games.

Those aren’t exactly stressful on the CPU then. That’s sort of the catch that Shifty is getting at. We don’t need this incredible amount of CPU to do just that. You just need to form the behaviours and the right behaviours to constitute fun for the player

 

[iroboto]

There's obviously scope for smarter challenges as shown in the Souls games et al, but I reiterate that most game are going to be more fun for most gamers with rather dumb AI for the reasons Esrever states.

I never found souls games to be smart as in reacting to the environment or a method of trapping or working together. Just the same patterns repeated over and over again. And hoping the player figures out the trick to survive the pattern.

It’s basically an old take on how they extended the duration of the game, don’t make it smarter just make it harder so that you have to keep replaying the section over and over again through trial and error.

 

[Shifty Geezer]

I never found souls games to be smart as in reacting to the environment or a method of trapping or working together. Just the same patterns repeated over and over again. And hoping the player figures out the trick to survive the pattern.

Sorry, I meant in terms of challenge, not AI. Most gamers want gameplay they can win at and win repeatedly with only the occasional failure, whereas some are happy to die over and over until they get good.

 

[Michellstar]

On the subject of how to get smarter intelligence though, the obvious route these days would be machine learning. Send huge amounts of gameplay data from real human players to The Cloud and have it churn out strategies that can be folded into a simulation. I've no idea how that would be implemented though!

Wasn´t this what Turn 10 promised for The Forza series?
What was it called? Drivatar?

For those who played it, It´s still a thing in Forza 7?

 

[Tkumpathenurple]

Intelligent artificial intelligence might not be conducive to a fun game, but more aware artificial intelligence might. At least, I should imagine that it would be more engaging.

Imagine you get into a bar fight in a Red Dead game. The standards we've seen this generation would result in you and the NPC getting into a punch up, maybe with some reinforcements spawning.

A more intelligent set of NPC's would result in the enemy grabbing bottles and glasses from the bar, breaking chairs over your character, eventually making enough of a hubbub to garner the attention of nearby NPC's. A crowd could gather and the occasional bystander might get hit accidentally, causing them to enter into the fray. A different bystander recognises the last one, and remembers him for a wrongdoing from time ago, so seizes the opportunity to leather him. That domino effect carries on for a little while, and before long, you have a full scale brawl.

Something like the nemesis system, but between NPC's. "You slept with my wife/ shot my brother/ took my job" type of stuff. Would that be computationally expensive, or is such a system just a matter of design?

Btw, I've yet to play Shadow of War, so if this is already present there... well, that's that settled then.

 

[milk]

Those aren’t exactly stressful on the CPU then. That’s sort of the catch that Shifty is getting at. We don’t need this incredible amount of CPU to do just that. You just need to form the behaviours and the right behaviours to constitute fun for the player

Well, imagine for exemple, each npc having multiple dynamic "thermometers" for diferent emotions like stress, fear, rage, etc, being affected by how agressively the player plays, tracking and evaluating at realtime the players play-style, and finding patterns in it and interpreting it can get quite cpu/memory intensive depending on how soohisticated the dev wants to get with this stuff.

 

[Shifty Geezer]

Something like the nemesis system, but between NPC's. "You slept with my wife/ shot my brother/ took my job" type of stuff. Would that be computationally expensive, or is such a system just a matter of design?

You don't need any of that background to produce the same results. The player has no idea who shot who's brother or took whoever's job. All you'd need is some simulated behaviour. RNG bystanders for whether they run away or get involved, weighted by a personality. Actually in GTA you wouldn't even bother with that because the personality itself would be RNG. You only need a personality if encounters are repeated and you want consistent behaviour from the same agents.

The thing that would be challenging is interactive environments and environmental awareness. This spatial processing strikes me as the most computationally expensive part. You need the whole environment structured such that searches for local interactives are possible. At first glance this may tie in with a global illumination system. If we're already partitioning the scene for lighting, that partitioning could be used for AI searches too, suggesting a more generalised advanced in scene representation than brute-force object position vectors is something that should be pursed.

 

[Shifty Geezer]

Well, imagine for exemple, each npc having multiple dynamic "thermometers" for diferent emotions like stress, fear, rage, etc, being affected by how agressively the player plays, tracking and evaluating at realtime the players play-style, and finding patterns in it and interpreting it can get quite cpu/memory intensive depending on how soohisticated the dev wants to get with this stuff.

Not in my limited experience. You have a bunch of values and do a bunch of maths on them every few frames. Even every frame, you're talking flops, not gigaflops. What I see making up much AI in my personality based projects are endless proximity tests, ray (or circle) casts, angle calculations, and physics look-ups. eg. I have an enemy space ship. It wants to know the angle to the player to know if it should flee or attack, in simple terms. You then project a motion vector to see if it'll hit anything on it's current trajectory and needs to evade. But for truly realistic flight, you need to test nearby asteroids for intersect vectors well in advance. So basic flight consists of proximity tests for local physics bodies, intersect tests based on motion vectors, angle tests to determine which side to fly to if necessary, factoring in acceleration and turn rate...lots of maths to simulate what the player does so naturally in seeing an object moving kinda in their direction and moving out its way, even luring an enemy into its path so they get walloped instead. And that intelligence advantage from the player is I think much of the appeal of many games. In Star Control, small ships could totally overpower big ships with good play, like getting into an orbit around the planet. There was challenge in being able to pull off the manoeuvres which possibly would have been reduced to frustration if the computer could properly adapt. It's have been impossible to beat an Ur-Quan with a Shofixti if the computer wasn't dumb, but that David v Goliath victory was always satisfying.

 

[Esrever]

Intelligent artificial intelligence might not be conducive to a fun game, but more aware artificial intelligence might. At least, I should imagine that it would be more engaging.

Imagine you get into a bar fight in a Red Dead game. The standards we've seen this generation would result in you and the NPC getting into a punch up, maybe with some reinforcements spawning.

A more intelligent set of NPC's would result in the enemy grabbing bottles and glasses from the bar, breaking chairs over your character, eventually making enough of a hubbub to garner the attention of nearby NPC's. A crowd could gather and the occasional bystander might get hit accidentally, causing them to enter into the fray. A different bystander recognises the last one, and remembers him for a wrongdoing from time ago, so seizes the opportunity to leather him. That domino effect carries on for a little while, and before long, you have a full scale brawl.

Something like the nemesis system, but between NPC's. "You slept with my wife/ shot my brother/ took my job" type of stuff. Would that be computationally expensive, or is such a system just a matter of design?

Btw, I've yet to play Shadow of War, so if this is already present there... well, that's that settled then.

Although stuff like this would be cool, it's nothing really intelligent about it, it just has to be scripted. There is no reason why games can't have this now and I think some of them do have mechanics which are similar but in the end it comes down to how much time the game developer wants to script scenarios that they want to pop up in their game. Nintendo has put a lot of small cool and interesting things into games like Super Mario Odyssey and BotW without the need of super AI processing or anything. In the end this comes down to attention to detail and how much each NPC/object can interact with others.

AAA games today lack a lot of the attention to detail because they are made to look pretty and sell based on hype, studios are just prioritizing what to spend dev time on and scripting object interactions for a lot of different cases just isn't that high on the list. Also each interaction requires a lot more time due to how much more time animations and modeling takes to make nowadays in AAA games. I don't think more advances in software or hardware will change this.

 

[iroboto]

Well, imagine for exemple, each npc having multiple dynamic "thermometers" for diferent emotions like stress, fear, rage, etc, being affected by how agressively the player plays, tracking and evaluating at realtime the players play-style, and finding patterns in it and interpreting it can get quite cpu/memory intensive depending on how soohisticated the dev wants to get with this stuff.

I think Shifty and Esrever covered most of it.

The only time you'd want say.. machine learning style AI algorithms is if you want a completely procedurally generated video game, which would require procedurally generated events and AI. Then the idea of that makes sense.

But if your game is linear, you're saving a lot of cycles by scripting everything. It's a lot of man power effort as the game gets larger (see Bethseda titles) but it runs on weaker systems.

If you want procedurally generated AI, you need a lot of power (or like some tech like custom tensor cores) and you could have a lot of procedurally generated AI the realtime rate we need it at. But then you get into a whole other mess like, fine tuning the experience since every time you try to debug the AI you're going to get different results. See procedurally generated game demos of No Man's Sky where in 1 demo they show X planet but upon restarting the demo cannot show you that planet again.

So there are problems whichever way you go. But because people are looking for a movie/cinematic style experience then having scripting through man power makes a lot more sense. Thus we've created a movie.

When we want procedurally generated AI, like we are trying to make Westworld the game, then that makes sense.

Today most procedurally generated AI plays abstract games with a finite set of rules, ie Chess, Go, strategy games.

Once story is involved.... ehhhh tough to demo that which cannot be repeated.. Everyone's experience is going to be different.

 

[iroboto]

Wasn´t this what Turn 10 promised for The Forza series?
What was it called? Drivatar?

For those who played it, It´s still a thing in Forza 7?

Needs a lot of work. They need to fine tune it. Remove other players entirely. And just focus on different classes of players driving. And not trying to play destruction derby.

 

[milk]

I think Shifty and Esrever covered most of it.

The only time you'd want say.. machine learning style AI algorithms is if you want a completely procedurally generated video game, which would require procedurally generated events and AI. Then the idea of that makes sense.

But if your game is linear, you're saving a lot of cycles by scripting everything. It's a lot of man power effort as the game gets larger (see Bethseda titles) but it runs on weaker systems.

If you want procedurally generated AI, you need a lot of power (or like some tech like custom tensor cores) and you could have a lot of procedurally generated AI the realtime rate we need it at. But then you get into a whole other mess like, fine tuning the experience since every time you try to debug the AI you're going to get different results. See procedurally generated game demos of No Man's Sky where in 1 demo they show X planet but upon restarting the demo cannot show you that planet again.

So there are problems whichever way you go. But because people are looking for a movie/cinematic style experience then having scripting through man power makes a lot more sense. Thus we've created a movie.

When we want procedurally generated AI, like we are trying to make Westworld the game, then that makes sense.

Today most procedurally generated AI plays abstract games with a finite set of rules, ie Chess, Go, strategy games.

Once story is involved.... ehhhh tough to demo that which cannot be repeated.. Everyone's experience is going to be different.

Thats why havung a good Director AI is also important, coordinating the actions of the npcs to keep the gameplay within sertaun constraints. L4D took a small step in that direction, but there are thousands of ways you could evolve that concept.

 

[iroboto]

Thats why havung a good Director AI is also important, coordinating the actions of the npcs to keep the gameplay within sertaun constraints. L4D took a small step in that direction, but there are thousands of ways you could evolve that concept.

yea it's certainly a talent to be able to make AI both challenging, and not requiring the enemies to be bullet sponges and simultaneously not make it feel like the AI is cheating

 

[Orion]

I think it would be interesting if you could create a realistic sims like game which had such good ai it appeared as if it were control of real humans. With voice synthesis and natural speech patterns.

 

[Deleted member 11852]

The only time you'd want say.. machine learning style AI algorithms is if you want a completely procedurally generated video game, which would require procedurally generated events and AI. Then the idea of that makes sense.

Not only those. Took at how the enemy AI adapts to the player's tactics in MGS5. If your main tactic is to sneak into camps and and pick off lone soldiers one by one, the camps adapt and start deploying soldiers in teams of 2 of more. If you're sniping from range a lot, they begin to wear body armour. There's a lot of responses to the various play styles.

 

[rekator]

Maybe they had to make cutbacks due to how the networking evolved?

Or make cutbacks due to how gamers evolved? ;-)

Don't think that gamers really like to fight AI with "Halo or Gears type" in a game where gamers like redo many times, and the more quickly, the fight to have more loot in a short time…

 

[iroboto]

Not only those. Took at how the enemy AI adapts to the player's tactics in MGS5. If your main tactic is to sneak into camps and and pick off lone soldiers one by one, the camps adapt and start deploying soldiers in teams of 2 of more. If you're sniping from range a lot, they begin to wear body armour. There's a lot of responses to the various play styles.

That’s still a state machine. It’s just closer to what Civilizations does per race. It’s a larger behaviour table and still does not require huge amounts of real-time AI. It can take time to process the AI before you start the mission, but it is not performing real time adaption.

 

[Deleted member 11852]

That’s still a state machine. It’s just closer to what Civilizations does per race. It’s a larger behaviour table and still does not require huge amounts of real-time AI. It can take time to process the AI before you start the mission, but it is not performing real time adaption.

Units are obviously not magically spawning body armour, the changes happen between missions so as to give the appearance of the enemy adapting to the player's tactics. Civ doesn't make much use of fixed data trees for anything much, it's why the AI is infuriatingly unpredictable when it ought to be based on the core traits of any given civilisation.

 

[Shifty Geezer]

Not only those. Took at how the enemy AI adapts to the player's tactics in MGS5. If your main tactic is to sneak into camps and and pick off lone soldiers one by one, the camps adapt and start deploying soldiers in teams of 2 of more. If you're sniping from range a lot, they begin to wear body armour. There's a lot of responses to the various play styles.

At the same time, that could be frustrating for gamers who like to snipe from afar and don't want to be forced into a different play style. Almost as if progressive challenges needs to be an option, leaving the option to cheese enemies for the more casual player.