Tobold's Blog
Wednesday, April 20, 2016
Artificial intelligence and real emotion

The graphics card in my PC has more computing power than the mainframe that controlled the first moon landing. The latest games have such evolved graphics capabilities that faces become realistic enough to not cause an uncanny valley effect. We are close to virtual reality graphics becoming mainstream. A huge percentage of the development budget of games over the last 30 years has been attributed to better and better graphics, and the result is showing.

On the other hand my PC doesn't have an "AI card". The science of AI has evolved, with computers now being able to beat human grand masters at highly complex games like chess or go, or master much less formal tasks like answering Jeopardy questions. But the artificial intelligence in most computer games hasn't evolved much at all. Computer opponents frequently are mentioned in video games reviews for their extreme stupidity in cases where that stupidity is so obvious that it breaks immersion. After 15 years of development of AI in the Total War series of games, the AI hasn't become better at playing that game. Video game advertising will frequently mention graphics, but almost never AI. It doesn't appear as if AI development is a major part of the development budget of any new game.

Instead developers are increasingly relying on other players to provide intelligence. It appears to be far cheaper to make your game PvP than to create a half decent AI. Why bother creating NPCs which behave believably in a MMORPG, if you can develop a MOBA instead where the NPCs are by design extremely stupid, and any opponent intelligence is provided by another player? Any game which ran into trouble during development is released as PvP only, because apparently the graphics was all the team worked on for the major part of the development process, and slapping on an AI at the end was considered optional.

The problem of that is that the PvP is only inherently attractive to a part of the player base, the "killers" of the Bartle test classification. A lot of players would rather either interact peacefully with others, or interact with a virtual world powered by artificial intelligence. If you consider other players as a cheap replacement of artificial intelligence, you risk getting more than you bargained for: Players not only come with intelligence and believable behavior, they also come with real emotions. And in  a conflict-based environment those emotions can run high and become rather unpleasant. The above mentioned MOBA games are frequently mentioned for having "toxic communities". In order to contain the toxicity, game developers then try to limit communication between players, which further drives away social player.

I am currently enjoying Magic Duels very much, because it has a decent AI and I don't need to play against other players. On the one side that allows me to avoid typical interactions with angry opponents, like people playing deliberately slow to annoy you. On the other side I do not need to be careful not to hurt the feelings of my opponent, for example I can throw games when in spite of a mulligan I still haven't got enough mana or get mana swamped. And I don't get the feeling that, compared for example with Hearthstone's human opponents, I am really missing out on much if I play an AI instead of a human. You really don't need all that much of an AI before playing against that AI becomes better than the sum of the advantages and disadvantages of playing against a human. How often do people make friends with their opponents in a PvP game?

When the internet and online gaming was young, many people believed in huge opportunities for friendly online social interaction through games. We believed we would get virtual worlds which we would inhabit online, build communities, and find friends. Today "multiplayer online PvP" is often a crutch for game companies too cheap to develop a decent artificial intelligence, and the internet is full of horror stories of online social interactions gone wrong. If a friend is somebody who helps you move your furniture, and a true friend is somebody who helps you move the body, how many friends that qualify did we really make? Didn't we just cheapen the word "friend" by sticking it to any online acquaintance that wasn't horrible to play a game with?

I do believe that today the "online multiplayer" feature drives away at least as many potential customers of a game as it attracts. Time to start working on artificial intelligence instead. We can still hope for great virtual worlds to live in, but there better be a lot of AI-controlled NPCs to interact with to make that a pleasant experience.

I would love to have a good AI to game against. For this AI to be fun, it should have the same strenght and weakness tahn human or it would be really horrible. In a FPS having a AI that one-shot(= perfect aim) you is very easy, but making smart decision you can predict is harder. In the Go game, even when the AI win, some movement was not understood by the commenter - and thus not very fun to battle. And simulating human error could be harder than simulating human smartness.

Still any step in this direction would be interesting.
In the Go game, even when the AI win, some movement was not understood by the commenter - and thus not very fun to battle.

I only partly agree with that. I am aware that in the Go game the computer made moves that were "not human moves", that is moves that no human in the history of the game had come up with as a strategy, and that somewhat destabilized the human player. But those weren't errors, but good moves. Just like in the history of chess sometimes people did come up with new moves, that is ultimately a gain for the game.

The problem in a FPS is that aiming for a human is not an intellectual exercise but a question of eye-hand coordination and speed. A good AI would have to somehow simulate the difficulty of aiming quickly that a human has, not just the intellectual exercise of deciding when to duck and when to shoot.
I think for any AI to be really believable it would need free will. Can't see that happening anytime soon, and if it does...

I for one welcome our AI overlords.
I agree.
I was so excited for the StoryBooks part of EQN.

I always thought something like ED or EVE could really benefit; the universe would be alive with ships 24x7 even if most were not human. AI is not just for NPCs; I was always curious what if your character was in the world 24x7, even if you were logged off. (Silly and poor environmental that DBO automates running around on your horse but makes your client be logged on to do so.)

I think part of the problem is so many game devs got there through other routes than computer science. I.e. another case where the games would be better if there were fewer elite gamers and more market-driven business types running development.

tl;dr: More AI? yes please.
The PvE encounters are "grinding", slaying large number of monsters. It's even true for EVE Online. It's a design decision, not a flaw: it's easy to make a hard encounter against one NPC with current AI, see mythic boss fights, yet developers stick to very stupid NPCs. I believe good AI is not made for two reasons:
- Bad players could simply not progress if the wolves in Elwyn Forest killed them by interrupting their spells or deny them kills by running away (any intelligent creature is fleeing from a stronger enemy)
- Worse than AI players would be unwelcomed in groups: why take an AFK-ing, chatlolling punk if an AI does better DPS.

In short: the bad AI is made so there is someone in the game who is dumber than Arthasdklol. And to do that, the AI needs to be VERY bad.
Gevlon is mostly right. As far as I understand it, there has never been any significant technical barrier to having much less predictable or more "realistic" AI in games. I recall reading interviews with developers around the turn of the millennium where it was claimed that they already had very much more sophisticated AI systems available than those they were using.

The issue has always been that players, by and large, want to win. The better the AI, the harder that is. Plus, of course, players tend to see good AI as unfair: the AI, after all, has the potential to know far, far more about any situation in the game than the player. Consequently AI has to be less able than the target customer base for the game and for any game developer that wants to serve anything other than a very limited niche market that means a low bar.

I think the place for strong AI is going to be in virtual worlds that are not primarily games. If the customer is, basically, the audience rather than an actor in the virtual experience than the more sophisticated the AI the better. In games, though, I think we will be seeing very dim NPCs and monsters for a long time yet.
I believe the appeal of PvP for most PvP players requires a great deal of not seeing their own bias. The half the time they win, they are brilliant and skilled. The half the time they lose, there is always some excuse, usually about incompetent teammates. For those players, they can feel superior to their opponent 90% of the time, the same as a PvE player.

You are capable of seeing through this kind of bias, and it isn't in your nature to throw blame on others, so PvP doesn't appeal to you.
@ Gevlon

That could make sense for WoW but doesn't make any sense for many other online/offline videogames where AI hase never-ever been really "smart" or "challenging" as Tobold suggests, even if the developers used words such as "amazing AI, smart enemies, etc".

There are many games where a great AI would make a game greater. That's still not something we can see, not yet. The tech isn't available and I am sure it will take a lot of time to se a human-lime AI opponent.

I think there are different types of difficulties. With WoW, the AI of the average mob could stay unchanged, with the difficulty simply raised by upping a mob's hitpoints and attack power. But this not the type of AI difficulty Tobold is talking about. Chess AI is not challenging because its pieces are more powerful, but due to a high skill level.

For example, the "Insane AI" in Starcraft is only better because the computer cheats by outproducing the player.
@Eaten by a Grue
Gevlon specified wolves using tactics like interrupting players. While humanoid mobs WILL try to run away, they wait too long to be a real threat. A "smart" AI would interrupt your cast, stun you, then run away and get 10 of it's friends.

The problem with this kind of AI is that players will simply game it by ganging up on the mobs, thus reducing the difficulty to trivial, and being angry about being forced to group at the same time.

At the same time, you cannot "tune" this kind of AI against an individual (assuming the mob isn't going to try to run and get friends.) because different classes have different ability sets.

that leaves you with lowest common denominator mobs who are dumb as a bag of hammers.
Ah, I think murlocs do this a little bit.
The reliance on PVP instead of content is what brought me to finally uninstalling The Division last week. I had really enjoyed the levelling process and was looking forward to some endgame content, but virtually everything pushes players to the PVP area.
AI is a hygiene factor. Graphics are motivators. Poor AI makes a game unplayable, but once the AI is good enough there is little incentive to make it better. If you face a strong AI at best it will go unnoticed and at worst you will think it cheating.
I know what I'm going to say kinda screws single-player games, but hear me out. I don't think that the way to go is more AI, but rather more social interaction.

You mention briefly PvP, but co-operative PvE gameplay can be as much enticing, if not more. For example way back I played on a vanilla WoW PvE server. The AI was as stupid as it gets (obviously), but the game was built in such a way that it promoted social interaction at every turn. In the end, even without PvP the server population provided all the necessary dynamics for an engaging experience.

The current problem stems that nowdays, most developers consider PvP the only viable social interaction you can get in a game, which leads to more people seeking out AI-enabled gameplay more and more.

I think devs exaggerate their ability to make good AIs, because it's often terrible where good AI is desirable. Look at Civ, for example.

That said, examples have been given of where artificial stupidity is required, which indeed includes most mobs in RPGs.
You know, I really want to see a co-operative PvE game where the AI is done right. I think PvP is for MOBA games where there is no leveling and you just jump in... but even then, I have no interest because latency and connection variances are too prominent.

So, at the line level, like wolves in Elwyn Forest, the AI has to be stupid. The simple reason for this is that there are a LOT of wolves there, and there is no reason they would not simply form packs and slaughter solo or small groups of players. That AI has to be strategic. Wolves engage in single combat, perhaps wait until they are nearly dead to run, etc. Really stupid AI.

Small camps of bandits are the same thing. It's absurd to think they won't hear you coming... they have the home turf advantage after all. If you walk up and see 4 guys in a bandit camp, and they actually had good AI, then you should know that the fifth bandit is right behind you about to stab you in the kidney. So they have to be pretty stupid as well.

So... put all that in the game. Those are the line NPCs you grind. But you also put in solo Apex Predators that don't assist the line NPCs, and aren't assisted by them. In fact, the line NPCs will refuse to engage the Apex Predators out of fear. Historically (Think the Griffon or Giants in Everquest.) this was done by having NPCs at a much higher level roaming about. But these mobs could just have much better AI. They interrupt your spell cast, they stun you. They fear you into line NPCs that then won't interact with the Apex Predator. They're brutal. Their only drawback? They're solo.

Instead of screwing around with trying to balance PvP, work on the Apex Predator AI. The Apex Predators are the PvP equivalent of gankers, but without the tear harvesting. The Apex Predators ARE beatable, they're not level 40 when you're level 10... they have superior tactics and response time, but their weakness is they have no friends. If you bring too big of a hunting party... they'll disappear. But if you know their pattern and set a trap, you MIGHT be able to bag one.

But in any case, you will fear the forest.
Your PC actually has an AI card. Since the graphics market plateaued NVIDIA is betting big on AI and repurposing GPUs to do the heavy lifting in AI. Now we have just to wait for the rest of the game industry to catch up. Sadly all the cool creative AI advancements are happening outside the game industry.
Also just to be clear (sorry can't edit previous post) game AI is more than just pathfinding or NPCs. Things like procedural generation of levels, texture generation, animations are all tasks that can be revolutionized by AI - and they will.
First off. Really good post tobold.

For some of the commentors:

There is a difference between good AI and "strong or difficult" AI that might be overlooked because they often overlap. Remember there are bad players too. A good AI is one that is hard to recognize as artificial, a strong one is one that whups your ass. Wolves are animals, and even though it is true that the wow example wolves are "bad ai" making it better does not neccesitate making the fights a lot harder. Tweak the numbers if you please, so that the small hunting pack of wolves are a fair fight for the adventurer. Make the wolves hungry enough that flight is off the table (if you think that is an issue), or have it flee if you like that better. It likely wont flee right away, and it might be wounded.
Mobs in WoW vanilla used to flee a lot, and it used to cause a lot of trouble as well. They also called for friends, etc. (Clearly scripted ehen though). Such AI could be made better, as in less obvioisly scripted/more realistic, without neccesarily becoming a lot harder. Give the player tools to negate the actions (silence = cant cal for help, root = cant flee), and give the mob a "reason" not to just avoid the player completely from the get go. Defending its home/predatory/stupid (yes you can have a good stupid AI if you want).

Tldr. Dont forget the difference between good AI and strong AI.
A good AI is one that can choose the optimal move from those available to it. The AIs in games like Civ and Paradox games are indeed bad, because even though in terms of rules they are on equal footing with players, they make poor decisions so they are easy to beat. Thus they need to "cheat" (i.e. have information or moves not available to human players, or other kind of advantage) in order to be competitive.

Now making a WoW mob do stuff like refusing to fight or fetching allies is not giving the mob a better AI. You could give the mob those moves without giving them the intelligence to use them effectively (for example, always trying to fetch more allies even though the fight is already skewed in favor of the mob, or trying to do it too late to matter, etc). And more importantly, this would alter the rules of the encounter, which may not be desired. Continuing with the example, trying to catch mobs all the time would probably get boring quite fast.
I also think online games are poisoned by the "Us vs Them" basic design that pits one group of players against another. In WoW, it's Alliance vs Horde, in dark Age of Camelot it was "RvR" with the three realms in constant combat.

You can't develop a "social cohesion" to your online "home" when all the lore makes it clear you're a Murder Hobo who is expected to "hate" the other faction. Further, your "home" is just a facade. All the good gear is spontaneously generated by boss mobs who live in some raid. Where did they get that? How can you have any respect for your own civilization when you have to get your gear from some deus ex machina mechanism?

At the same time, anarchy never works, so just giving players house deeds and letting them go nuts with them does not work.

What I want to see: A town is designed and already laid out, but there are no buildings in it. You can buy a building permit and the requisite materials, then find or buy a vacant, pre-zoned plot for it and build. You have a degree of ownership in the town, and all commerce and technology for your civilization is there. Of course, all players are part of that civilization, the "bad guys" are NPCs.

To get raw materials to build, you find them in the encounter areas outside of town. To get materials for advanced weapons and gear, you beat encounter bosses (Raid, scenario, solo... whatever.) and bring them to town where a craftsman who runs a craft shop makes the item you want for a price. There should also be a "craft commons" where players that do not own stores can craft items. They could then sell the items through a store via a brokered trade mechanic that the store itself has little control over.

A lot of these concepts have been tried, but there have always been fatal (Clearly visible in hindsight.) flaws to the concepts. Instead of fixing the problems and refining the cooperative multi player model, the PvP model has been the default.

The biggest flaw, in my view, is the assumption that anarchy would work. You need to pre-zone the town, and have a brokered trade mechanic for the buying and selling of goods. Your gear needs to come from YOUR civilization. Gear you then use to fight the darkness encroaching on YOUR town.
People think anarchy will work because they have a casino mindset - they think they are going to win big. And when it doesn't work and they lose, they blame the game and move off to the 'next big win'.

Some of what you're talking about has been tried, but only in open-PvP games as far as I have seen. The kind of players who would thrive in a community building game (and the vast majority of all players, really) are simply unwilling to play a game with that kind of environment.

The other issue is that these systems usually turn into a wasteland of abandoned buildings as far as the eye can see. A lot of this is because the buildings are entirely cosmetic. You build it, look at it, and then have no reason to return because it is totally useless. The other major problem is that games of this type try to make this a feature to give every last player. When you have hundreds of thousands of players, that just isn't going to work.

So I want useful buildings, that are prohibitively expensive for the vast majority of individual players, with a significant ongoing lease payment for the land. If you stop paying (usually because you quit the game), you lose it (getting rid of abandoned buildings). More casual players can team up with a guild, and now you have much more useful guilds, for instance a blacksmithing guild who has spent a lot of time building and upgrading a top quality forge.

Exactly, not everyone can have a "store in town." Most people need to use the common crafting area, which is always there. The price for a "store in town" tracks demand for real estate. You have a town because you need a town. If the town gets too populated, you spin off a new town and make the deeds available. Only as many towns as the game needs are ever made.

The only way a "building for everyone" works is with housing, and then it has to be instanced, with all the thousands of houses taking up the same amount of virtual real estate. I think Wildstar does something like that, with the house itself on a floating island in the sky. Dark age of Camalot has a "housing zone", but the houses in the zone were tiny placeholders. The actual house was a mile up in the sky.

All in all? I think to most important thing to change to make a "community oriented" game is to eliminate PvP.

I get the impression that Tobold was talking about in-town interaction AI as opposed to combat AI, but I don't think that's really needed... instead, you need an environment that fosters respect for the game's lore and civilization. It's YOUR town, not that place you have to go to to get flasks and use the forge.

Community oriented players do not want to "police themselves" in any meaningful way as it becomes immediately clear that doing so descends to anarchy. You cannot "police yourself" against murder hobos that just respawn and come right back. Likewise, any scheme to "allow that" and at the same time "keep the town safe" through the actions of AI are doomed to failure. The attempts waste development time and just result in a WORSE murder hobo problem. Likewise, any scheme that hugely penalizes PvP oriented players, like making it so if they have attacked a player in the last month they can't go into town or do group content, will just piss off those players. They should be told from minute one that a MOBA game is for them.

Developers can then spend the time they would be wasting with PvP issues to develop better group content (Dungeons, Raids.) and solo content (Quests, Scenarios, etc.)
@ Tithian

> more social interaction

That also can lead to weird and undesirable effects, just look at any "Looking for group" session or random-people raids: they often end in drama, kicks, etc. Interacting with humans doesn't always grant fun and/or great results, even in co-op. Plus humans often have the bad habit of being jerks and act like idiots if they're online.
Funny you should post about this now. Rulolfo Rosini mentioned this above, but Nvidia had a major announcement on this topic just recently. They have created a processor specifically targeting the "deep learning" AI domain:

Nvidia's CEO is quoted as being "all-in on AI". Newsweek covers this here:

Interesting times.

Post a Comment

Links to this post:

Create a Link

<< Home
Newer›  ‹Older

  Powered by Blogger   Free Page Rank Tool