Complaining Of Ai

[Belceburro]

One of the things that annoy me in all the thief-like games I played before was the lack of peripheral vision of the AI. Humans develop peripheral vision in the first stages of life, which allows us to see high movement in the corner of eyes, just like a reaction to an agression coming from the flank. In fact, we can detect more easily the movement at the corner of eyes than in the center.

I've never seen AI reacting like this and it takes out a lot of realism from my point of view.

I suggest increasing the chance of triggering AI when the thief is moving and the line of view to the AI is lateral or nearly lateral.

It should get a couple of lines of coding and i guess it can be easily added to the triggering routine you already got without spending lots of CPU.

 

Again, I'm surprised of your really nice work. Keep on coding this mod. You're great!!

[Ishtvan]

Thanks for the support and suggestion. Did you play T2X? The AI had pretty good peripheral vision in T2X. I'm not sure I liked the effect it had on gameplay.

[New Horizon]

Thanks for the support and suggestion. Did you play T2X? The AI had pretty good peripheral vision in T2X. I'm not sure I liked the effect it had on gameplay.

 

Indeed, that's something that worries me about a lot of the suggestions I hear. People don't seem to realize the implications a lot of 'great' ideas will have on the basic gameplay. Something that seems like it should be an improvement, could just end up making the game tedius, instead of fun.

[obscurus]

If you want to be realistic about it, it is actually easier to sneak up on someone in their direct line of sight than if they can see you out of the corner of their eye (obviously it is better if you sneak up from behind, so they can't see you at all). This is due to the way the eye is structured - there are lots of movement senitive rods around the periphery of the eye, but the fovea has mostly cones, which are relatively crap at detecting movement. They are very good at picking out patterns though, so if you are standing still, the guard will be more likely to spot you front on than if you are moving.

 

I don't see how having a realistic vision model for AI would be at all undesirable from a gameplay perspective (provided it doesn't impose an unacceptable performance hit), on the contrary, it is better if the AI responds as closely as possible to the way real people do. I haven't played T2X yet, so I am not in a position to comment on their implementation and its subsequent effect on gameplay, but gameplay is very much in the eye of the beholder so I take '[realistic feature XYZ] is bad for gameplay' type comments with a very big lump of salt.

[New Horizon]

...but gameplay is very much in the eye of the beholder so I take '[realistic feature XYZ] is bad for gameplay' type comments with a very big lump of salt.

 

Well, if there is no way to actually 'defeat' the AI...then there is no game, and as such, this is bad gameplay.

 

I don't think the goal should be to completely simulate realism, but rather to create the illusion of realism and provide enough of a challenge to engage the player, not make the experience a burden. If something is going to be 'that' realistic...I might as well go out and start Thieving in real life. There has to be some unrealistic elements, otherwise...you're just going to end up with a very unbalanced experience.

[Ishtvan]

If you want to be realistic about it, it is actually easier to sneak up on someone in their direct line of sight than if they can see you out of the corner of their eye (obviously it is better if you sneak up from behind, so they can't see you at all). This is due to the way the eye is structured - there are lots of movement senitive rods around the periphery of the eye, but the fovea has mostly cones, which are relatively crap at detecting movement. They are very good at picking out patterns though, so if you are standing still, the guard will be more likely to spot you front on than if you are moving.

 

Couldn't you say that spotting someone sneaking up to you is a combination of movement detection and pattern recognition? Movement draws the eye, but once the eye is drawn, you look directly at the thing to make out the pattern, right?

 

I'm pretty sure we've talked about this before, but maybe it was in the private forums. The FOV abstraction, coupled with a movement modifier that makes you more visible when moving, is actually not that bad a model. Think about how a real person's eyes would be randomly scanning over the FOV physically allowed by their head. We assume that they're scanning randomly over the FOV, so any point within the FOV is sampled by both rods and cones, over time. If there's movement, the rods will probably pick it up first, then the eye will be drawn to it and the rods will be utilized to recognize the shape. Therefore, we can say that, taking the time average, the detection probability is approximately uniform within the physical FOV. Plus, from a gameplay perspective, you can predict the AI visual response based on which way the AI's head is turned with the FOV method.

 

If you want to get into the difference in response between rods and cones, you either have to assume that the AI stares straight ahead in a fixed direction relative to the way their head is facing, which IMO is a rather bad assumption, or you have to let the eye move independently of the head, and model the psychology of how a real person scans their eyes over the FOV limit imposed by their head. Personally I'd rather not try to mathematically model the psychology of eye movements for every single AI and every single alert state. Also, this would mean the AI's visual response would vary unpredictably due to the random eye movements. The player couldn't guess at the visual response based on the pattern of head turning, because the eyes move much faster than the head and could change position at any moment.

 

Believe it or not, it's more fun when a plan based on patient observation of the AI's behavior is rewarded, rather than being randomly caught because the AI's rods happen to flick over you when you're moving at the edge of their FOV. We could give our thief random coughing attacks that would give him away, too, wouldn't that be fun!

[Belceburro]

Ok, I'll correct myself : Being reallistic at expense of gameplay is a truly bad idea, but you can consider simulating this effect maybe increasing the sight angle of the AI (only in higher difficulty levels of course). If you plan creating difficulty levels, sure you can offer a more realistic approach at higher level, IMO.

Maybe it is me, but i bore when i sneak and see the lack of perception the AI has and how innatural excelent perception it has when is alerted. This dramatical change affects my gameplay experience, because it appears that there is no diffence in perception of the "drunken" AI and the "unaware".

 

Maybe the suggested approach will not be triggering the AI if you are moving at his lateral, maybe increasing chance of detection if you do it "near". Sure Garret has innatural ability of being cloacked but, i don't like the AI not spot me if I am at his side 30 cm breathing over his shoulder.

 

I suggest if you find that a "great idea" crushes gameplay, but still worth it, then consider making available to the higher level of difficulty only .

 

 

 

Another thing is that sneaking ability is related directly to the architecture of the map. If you are moving inside a room plenty of statues or columns, you can hide behind easily than just standing in the middle of an empty room.

Maybe the fast and easy way to improve gameplay and also realistic, is to increase chance of detection (and show it in the light gem) if you're not stick to a wall or object. I think this does not craps gameplay and improves the way a thief moves: sticking to walls .

 

 

Lastly... any one can tell me why you can kill a AI with one Headshot when it is unaware and you need 3 or 5 headshots when it is alerted? . Ok, ok i see realistic = not funny

Edited February 13, 2006 by Belceburro

[sparhawk]

Indeed, that's something that worries me about a lot of the suggestions I hear. People don't seem to realize the implications a lot of 'great' ideas will have on the basic gameplay. Something that seems like it should be an improvement, could just end up making the game tedius, instead of fun.

 

That's why they are the clients and we are the developers.

[Belceburro]

That's why they are the clients and we are the developers.

 

Hehe, that's sure is realistic

 

But that's the way it works, we are expectant of your amazing work and want to help providing new ideas.

Sure you can disregard it if you consider they're not good enough... in fact it is your game

Edited February 13, 2006 by Belceburro

[sparhawk]

My comment was more addressed to that poeple think up new features without considering the gameplay impact, not that you put forward new ideas becuase that is ok. We had a lot of discussions already and most ideas are not new to us. The major problem is, that some ideas sound cool on the blueprint, but don't work well if you put it into code because it can get tedious or disrupt the gamplay and take out the fun.

[Belceburro]

My comment was more addressed to that poeple think up new features without considering the gameplay impact, not that you put forward new ideas becuase that is ok. We had a lot of discussions already and most ideas are not new to us. The major problem is, that some ideas sound cool on the blueprint, but don't work well if you put it into code because it can get tedious or disrupt the gamplay and take out the fun.

 

I agree

[bob_arctor]

Gah I had a great post going, pressed something and it went!

 

"Photopic vision is the vision of the light-adapted eye; in many animals, color vision, mediated by cone cells."

 

"Scotopic vision is the monochromatic vision of the eye in dim light. Since cone cells are nonfunctional in low light, scotopic vision is produced exclusively through rod cells. Vision in normal light with functioning cone cells is photopic vision. Mesopic vision occurs in intermediate lighting conditions and is effectively a combination of scotopic and photopic vision."

 

Cones= mostly in the macula (yellow spot) then in fovea then foveola. If a guard used this in bright light he'd see you even when standing still: lots of colour and detail. Scanning and the brain fills in the gaps so we "see" good detailed vision despite small area and the blind spot of the optic nerve.

 

Rods= See movement, black and white, in the dark, low detail. It'd be hard to use rods to see a still thief. Movement would alert the guard more.

 

Peripheral vision= if a guard saw you run past with it he would jump and turn round, see movement, but no detail so unless he then saw you he would think "I'm jumpy, must be rats" because he didn't see a thief, he saw some kind of movement. Moving slowly would probably make him suspicious slowly then he'd turn and get you!

[Belceburro]

Peripheral vision= if a guard saw you run past with it he would jump and turn round, see movement, but no detail so unless he then saw you he would think "I'm jumpy, must be rats" because he didn't see a thief, he saw some kind of movement. Moving slowly would probably make him suspicious slowly then he'd turn and get you!

 

That's what i was saying when i first posted. We should decide if it damages gameplay, and if does not, if it shall be pointed only for the highest level of difficulty.

[Ishtvan]

I suggest if you find that a "great idea" crushes gameplay, but still worth it, then consider making available to the higher level of difficulty only

 

You have to realise that we're working for free with limited time resources. If we decided to go off and program every single "potential great idea" that "may" destroy gameplay, just to put it in a high difficulty level, we would never have a finished mod. We're going to release the source anyway, so if someone is really passionate about something, they can add it in themselves as a 3rd party mod.

[obscurus]

Well, if there is no way to actually 'defeat' the AI...then there is no game, and as such, this is bad gameplay.

 

I don't think the goal should be to completely simulate realism, but rather to create the illusion of realism and provide enough of a challenge to engage the player, not make the experience a burden. If something is going to be 'that' realistic...I might as well go out and start Thieving in real life. There has to be some unrealistic elements, otherwise...you're just going to end up with a very unbalanced experience.

 

 

It is a bit of a leap to assume that realisticly modeled sensory algorithms for AI will so badly unbalance the game as to make it unplayable. It won't at all. What it will do is make the player play in a way that is a bit more real-worldly. Instead of just relying on shadows, the player will have to make use of cover, such as hiding behind pillars and barrels and so on. Of course you will be able to defeat the AI, just as a real-life thief can defeat real life guards through stealth, it will just not be what you are used to from playing Thief. You make the assertions about balancing gameplay, without realising that what you find balanced, others might find horribly unbalanced, so my point that gameplay balance is in the perception of the player stands. Sure, you need to make some concessions to un-reality due to the technical limitations of the computer game medium, but having realistic sensory acuity for AI is not in need of any concessions at all. I am speaking for myself here obviously, but I am not interested in dumbed down AI, I want it to be as close to the real deal as is possible win the limited capabilities of a modern computer. A game that models the laws of physics and nature as closely as possible will automatically be as balanced as reality is. Where you can't model the laws of nature well, you need to make concessions in order to restore the balance, but sensory acuity of AI in no way requires this.

[Domarius]

I'm currently playing T2X and I like the extra difficulty presented by the AI's more realistic FOV. It's one more step to making them more like real human opponents instead of stupid NPCs.

[sparhawk]

It is a bit of a leap to assume that realisticly modeled sensory algorithms for AI will so badly unbalance the game as to make it unplayable.

 

It's also quite a leap of faith to assume that more realism will make automagically make a better game.

 

You make the assertions about balancing gameplay, without realising that what you find balanced, others might find horribly unbalanced, so my point that gameplay balance is in the perception of the player stands.

 

Everybody is free to create his own game. Obviously we are doing this for ourselve. It's not as if we have to appease everybody at all costs. So if somebody doesn't like what we consider as balance, then he can either change it for himself, make his own game or accept it. But we have a very good cut through the Thief playerbase, which is really surprising. If we had tried to gather such a homogen selection on purpose I doubt we would have managed it. As such I think that if our team likes it, most players also will like it, because pretty much every type of player is represented.

[bob_arctor]

From the Rope Arrow thread:

"The fact is people have been playing Thief since 1998 so they will accept some unreal elements as part of the game."

 

What people have to remember is that you can't actually sneak about like Garrett in real life. Well maybe about 3 ninjas 400 years ago in Japan could, you know what I mean.

But most burgalries are not creeping past guards, blackjacking them, or anything. It's nicking stuff while people are asleep, with a knife or gun if they wake up from you making noise.

 

Thief is fiction, it's a game. So it has to have game mechanics. In real life anyone trying to be Garrett would get caught really quickly, and that would be a dull game.

[Ishtvan]

True, a lot of real life thieving is probably just spending hours or days in research and preparation, getting to know a neighborhood and when people are out of the house, what alarms people have, estimating floor plans and vauable locations from the outside, who has a guard dog, etc. The research phase can't be translated so well into a game. Either that or it's just a crime of opportunity with a high chance of getting caught or having to shoot someone.

[New Horizon]

I'm currently playing T2X and I like the extra difficulty presented by the AI's more realistic FOV. It's one more step to making them more like real human opponents instead of stupid NPCs.

 

Where did you hear they changed the FOV in T2X? I was told that the AI were not altered from the original games.

[Domarius]

I've noticed it myself. I keep getting "first level alert" seen when I don't expect to.

Whoever told you that may not consider AI vision cones to be AI (artificial intelligence) per se.

[New Horizon]

I've noticed it myself. I keep getting "first level alert" seen when I don't expect to.

Whoever told you that may not consider AI vision cones to be AI (artificial intelligence) per se.

 

Not sure, but they said they didn't touch anything like that for T2X. I could have misunderstood. I wonder if it has anything to do with the higher than normal ambient light

levels in T2X?

[Ishtvan]

Yeah I noticed it on my own as well, was playing T2 shortly before T2X came out, and then in T2X, it was like all of a sudden I couldn't pull off the moving in their peripheral vision that I used to for T2. Definitely seemed like the FOV was set wider, but maybe it was just the high lightgem level.

[bob_arctor]

The lightgem is certainly different, Zaya (That right? I may have forgotten) lacked the sneaky training of Garrett so had a worse ability to hide.

Dunno about AI. Didn't notice it myself I don't think. /shrug

[OrbWeaver]

The lightgem is certainly different, Zaya (That right? I may have forgotten) lacked the sneaky training of Garrett so had a worse ability to hide.

 

The lightgem appears different, and it does not function at all in the first level. I have not heard of any differences in its behaviour though.

 

There are problems with ambient light in T2X - in some areas it is almost impossible to find an actual shadow, as opposed to a dark area that leaves you visible.