Hacker News Reader

Comment by madez | Hacker News Reader

I wonder if AI is where games can make use of many cores. If the game logic can not be reasonibly parallelized anymore, the unused computing capacity can be used to add "intelligent" agents into the game. They don't need to do revolutionary things, just being more complex, more "reasonable" and less predictable than the usual agents would be enough. I think that has the potential to make some games more interesting.

Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy.

Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples?

View on HN

> Imagine a game like Warcraft 3 where the units actually seem to have a personality and follow "their guts" in battles rather than following simple patterns like attacking the nearest enemy.

> Maybe I'm ignorant about the current state of the art and games already do that. Do you know examples?

It's a really tough trade-off between fun and realism. It's cool that your units have personality for a bit "Oh, look, those 2 dwarves are fleeing from the troll fight early!", but not fun once they disobey your intentions "That fucking troll has 3 hit points left. If my 2 dwarves didn't run away, it would be dead! Now it's destroying my base!"

One game I remember striking a good middle ground with this was the original Battle For Middle Earth. It changed animations, sounds and other aesthetic things based on AI, while the game remained the same mechanically. If your units ran up against a mountain troll, you would see the ones in front of it cowering and holding up shields, while units behind attack furiously, but every unit was really doing the same damage. Or if you left troops of Orcs just standing around too long, they'd break out into fights and kill other orcs, but the animation is done as a trick where as one orc model "dies", a new one spawns in the group. So you didn't really ever lose Orcs, even though there would be orc corpses strewn wherever you had been camping.

I think the main reason though is just that AI doesn't really sell. Graphics, physics and voice actors sell. You can blame a failure on AI, but very rarely is success attributed to good AI in a game.

Bad AI often winds up panned in a game's review scores, which have an impact on sales.

For instance, bad pathfinding in an RTS game can make the game almost unplayable. Halo is just about legendary at this point for teammate AI that is worse than worthless [1]. Any one of these things can destroy immersion at a critical point.

AI is to video games as IT is to a modern enterprise. If it's doing its job, nobody notices, but if it's bad, it sticks out like a sore thumb.

[1]: https://www.youtube.com/watch?v=vdbg8h3d3uM

Or, if it does its job really well, it gets accused of cheating.

My former company had a backgammon game with a really good AI. We constantly got accused of cheating, even though the dice-roll algorithm was identical (it didn't have a parameter to specify who the roll was for). What people missed is that a good backgammon player will position themselves to benefit from a wider set of possible dice outcomes.

I was under the impression that backgammon was a solved game (i.e. there is a known optimal move for every possible position), can I ask why you used an AI instead?

Probably because an AI that always plays optimally wouldn't be very fun to play against.

Cannot find it right now, but I saw a video of Civilization 's Sid Meier saying that about Civ's playtests.

> Bad AI often winds up panned in a game's review scores, which have an impact on sales.

This is what I said though. A game's failures can be blamed on bad AI. But once you reach the point of okay AI, you stop hearing about it. There's very little reason to put in whatever effort is required to make your AI great, because the only people who will take notice are hardcore players who put in tons of hours. And that's nice and all, but the game costs $60 whether you play it for 10 hours or 100. So why spend millions on a feature that's hard to make a trailer for, and is only noticeable to your most hardcore fans?

>Graphics, physics and voice actors sell.

This is hardly arguable, but personally, after some time playing I begin to recognize animation/sound/timing (i.e. behavioral) patterns of non-plot entities and that ruins the entire experience. That's what retains me on the storyline and prevents free exploration -- there is nothing really interesting besides nice graphics. I think there is a room to evolve from [hundreds of] programmed random encounters to purely independent world.

Actually, I'm fine with GTA3/GTASA-like graphics, like I was with roguelikes and physical book-and-dice quests, so idk how that applies to wide public.

ps. Another conclusion, it may appear too frustrating. We probably play games to win and achieve. If there is no end, no goal and no perfect winning strategy, that understanding alone may cause depression, like IRL.

Not necessarily, we play games for play sometimes. Because it is fun. Take Minecraft for example. No serious win-state (until fairly recently) but it still has tremendous appeal.

I remember that in Close Combat https://en.wikipedia.org/wiki/Close_Combat_(series) the units have some small independent thinking and refused to follow very stupid orders and wanted to take cover. It probably has a few simple rules, like an expert system.

As discussed in other comment, it was a pain to manage them, because they didn't follow all the (stupid/suicidal) orders.

>I think the main reason though is just that AI doesn't really sell. Graphics, physics and voice actors sell. You can blame a failure on AI, but very rarely is success attributed to good AI in a game.

The main reason that I avoided buying Civilization 6 is that the AI was notoriously bad. For any kind of turn-based strategy game, I want the AI to be superb.

I'll note that this actually makes the original point, since despite your avoidance of the game, Civ VI is estimated to have been the top grossing game on Steam in 2016. So, even in this case, the game has sold gangbusters despite lackluster* AI.

*I didn't intentionally rhyme lackluster with gangbuster. Rather these are the words that sprang initially to mind, and it's a kind of satisfyingly weird combo, so I'm leaving it in.

Source: https://galyonk.in/steam-sales-in-2016-def2a8ab15f2#.nxpnuko...

> I think the main reason though is just that AI doesn't really sell.

I'm not sure that's the case. Last of Us and MGS:V both had intelligent AI where they would change their behaviour based on what you've done in the past and work together as a team to flush you out of a hiding spot. This was widely covered and praised in reviews and pre-launch interviews.

There's a recent game called Hello Neighbour where the eponymous neighbor learns from the patterns you use to sneak into his house. The one video I saw of the game looked freaky. 10/10 would play.

I don't know about current games (I suspect they have enough to load the CPU with gfx/physics to devote much to AI), but state of the art AI software can definitely be many core. It's very useful for anything doing Monte Carlo simulations, tree search, planning.

Games haven't really been pushing CPUs very hard for the past few years. All the mucking about with overclocking has been purely a matter of culture recently, as you can play top tier games in 4K with a mid-range CPU, as long as it can handle enough PCIe lanes for your monster GPU setup.

Depends on the game, really. That may be true for some, but far from all.

Take a look at the game sections from the recent Kaby Lake reviews. Here's Anandtech's for example: http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-...

Most of them look pretty much like that, with very little difference until you start getting to really low-end or old chips. Games are mostly GPU limited these days and those that are CPU limited are mostly leaning on one or two cores and thus haven't really gained much from recent generations.

You're arguing that benchmarks of recent top of the line chips (which at most tend to have a few tens of percent difference) not showing large performance difference is evidence that games aren't CPU limited, whereas the same test shows drops in single-thread performance immediately cause large performance differences (see AMD chips).

If anything, this shows that games aren't strongly multi-threaded (which comes with severe development and debugging impact) or aren't made to scale beyond the highest end chips consumers are likely to have.

Lastly, consider that while the graphics effects and physics effects of a game can be scaled down, scaling down the AI in many instances would cause balancing issues.

Since when is a $168 i3 chip "top of the line"?

Imagine an RTS or something similar where you play with an AI and he learns your behaviours. He tries to help you. At first you deny what he does, he learns, do something else. A bit like the beast in black & white!

Game logic has been parallelized since about 2005-2010.

I was told (but didn't verify it myself, nor have any sources on this) that in Half Life 2 (2004), your enemies would "learn" from your attack patterns and adapt to it. At the same time, as to make the game enjoyable/chellenging and not frustratingly difficult, the agents would adjust their difficulty level to your skill. Take it as an anecdote.

That was actually Left 4 Dead.

https://en.wikipedia.org/wiki/Left_4_Dead#AI_Director

The director in L4D was had a very simple logic with inputs like number of alive special infecteds and number of zombies to control spawns. It's like seeing a health kit when your health is low.

And the director in L4D2 actually changed layouts of levels to suit difficulty.

Have to start somewhere, though.

IIRC it tried to be slightly more clever (might've been L4D2) in that it also kept track of the recent history of the game and attempted to balance the pace of the action.

If it thought players had had it easy for a while it might throw a larger waves of monsters and specifically deny spawning health for a few minutes.

There's a good presentation somewhere about how it would try to balance the rising action with periods of "stress" and "relief" over the course of a level.

I've played an unhealthy amount of HL2 and spent a while modding it, and I don't believe this is true. People say things like this a lot about games with decent AI. For example, the old quake 3 hoax about bots "learning" to become peaceful on a server that was left running for 4 years. It's fun to imagine the AI as more than just a set of scripts.

The Valve Developer Community has some okay documents about the AI [0], moddb has some nice articles too (no links offhand). And most of the source for combine guards is available in the Source SDK [1], you can see tons of if's and hard coded sequences of actions.

I've read that the AI was changed after the Orange Box release, so maybe there was some learning features to them originally. Or maybe those features do exist within HL2's private source files, though it doesn't "feel" like it.

[0] https://developer.valvesoftware.com/wiki/AI_Learning:_Combin...

[1] https://github.com/ValveSoftware/source-sdk-2013/blob/master...

I've heard a similar anecdote about a racing game (I think it might have been in Forza) where the developer had a AI where individual racers could hold grudges and that would affect how aggressively they treated you (bump someone in lap 1 and they'll be more aggressive in lap 3). Eventually they had to take it out for various reasons, but didn't advertise that fact for obvious reasons. When the game was finally released several reviewers brought up the feature as a plus, not realising that they were just reading into the actions of the standard AI.

As a side note, the writeup of F.E.A.R's AI is a fascinating read. By combining very simple goals and actions, they created AI that is still remembered as incredibly good. http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear....

In ARMA opponent AI is a parameter. When you have a lot of players on one server, you turn down enemy AI so you don't bog down the entire server with pathfinding calculations.

Interested in this as well. Applying machine learning in this way could create some interesting game play.

pretty sure Rival Theory was working on something similar to that

Mr Anderson...

Reply on news.ycombinator.com