Return to “General”

How superstitious should NPCs be?

Not at all -- NPCs doing dumb things would weaken the challenge of the game.
Total votes: 6 (8%)
A little bit -- it would help make the social world feel more plausible.
Total votes: 63 (81%)
Very superstitious -- it would be fun to see NPCs doing crazy things.
Total votes: 9 (12%)
Total votes: 78
Post

Very Superstitious

#1
I'd like to talk about superstition.

No, not that Superstition (although it's exceptionally cool). I mean the process of associating an action with an observed effect when there's no real causal connection between them.

Humans (some humans, anyway) do this all the time. We avoid walking under ladders or opening umbrellas inside a house or letting a black cat cross our path (to avoid bad luck); we pick up a four-leafed clover or a penny (to gain good luck); and if we're sports figures, we spit, rub ourselves, turn in circles, and perform all kinds of contortions and twitches out of the belief that these actions can change events in the world with which they are completely unconnected.

The reason why I'm wondering about this is because it occurred to me today that NPCs in LT -- if I've understood what I've read -- are going to have three very interesting capabilities:

1. NPCs can form plans to achieve goals. These plans are representations of beliefs about how the NPC's actions can produce specific desired changes in the world.

2. NPCs can perceive effects that occur in their world. These perceptions are new information about the world added to what the NPC currently believes to be true.

3. NPCs can modify their plans based on goal-seeking logic that takes into account world-effects that they've perceived. This creates a feedback loop of action, observation, thinking, and reaction.

Where superstition comes into play is when it is possible for an NPC to take an action, to perceive some world-effect that is in fact completely unrelated to the action they just took, and to mistakenly correlate the action and effect so that they change their plans -- and then their future actions -- based on that false correlation. At that point, the NPC is behaving superstitiously.

So if NPCs in Limit Theory have the same basic kind of action/effect/perception/assessment/re-planning/reaction loop that humans have, it's possible that they could form superstitious beliefs as humans do.

Suppose an NPC who focuses on trading leaves a space station. For whatever reason, he takes an action to spin his ship in a circle. This just happens to be followed immediately by the arrival of a merchant ship that's full of cargo they want to sell at a discount.

If that NPC can know they took an action (spin in a circle), and that a merchant ship arrived immediately after their action (allowing that NPC to satisfy his goal of making good trades), can that NPC ever reason that spinning in a circle causes well-stocked merchant ships to appear, and get into the habit of spinning in a circle every time he leaves a station to try to cause a merchant ship to appear?

If so, is that awesome? Or horrifying? And that's just one contrived example -- imagine if you can all the possible superstitious behaviors that might emerge in an LT universe where NPCs can observe and plan. :)

Some questions about this notion:

1. How many times does an action/effect pair have to happen before an NPC decides there's a real correlation between them? Once would allow a lot of superstitious behavior to emerge; five times would be a very suspicious universe in which NPCs take quite a long time to accept that one behavior is better than another. Assuming NPCs count action/effect pairs at all, should the trigger number be fixed for all NPCs? Different for NPCs based on their personality? Different based on the importance (including survival importance) of the action/effect belief?

2. The flip side of the previous question: how many times should a superstitious belief not work before an NPC changes his mind, decides there's no real correlation there, and stops performing that action? Should that depend on the NPC's personality, so that some NPCs keep any superstition they've adopted?

3. What is the scope of what NPCs can observe about the world? Can they only see direct consequences of their own actions, which would mean that the "causal connection" concept they form between their action and the subsequent world-effect is always real and correct? If so, then how can NPCs ever express strategic behaviors, which require seeing patterns of cause and effect across large stretches of both space and time?

4. If superstitious behavior can emerge, should it be considered a programming fail state? In other words, if some NPCs occasionally take less-than-optimal actions because they've formed a superstitious belief, is that a bug that needs to be corrected? Or is it a glorious emergent effect that makes the world of the game more enjoyable?
Last edited by Flatfingers on Thu Jan 30, 2014 3:29 pm, edited 2 times in total.
Post

Re: Very Superstitious

#3
i would like to see "suboptimal" behaviour emerge out of the AI,
I dont belive that AI in LT will be able to create "optimal" plan anyway. It is simply too complex for it to handle on perfect level, there are too many factors and many different NPCs that will change those factors with their actions, so no need to worry i think.
And its just the way how it should be.
I̲̩̳̺̩̫n̵̻̘͚͖̗͎ͅ ͢J̜̬̗̦o̩̘̦̪͕͉ͅs͞h̞͘ ̯̹͈͙w̯̙̥e̱͉ ̬̙̘̭̯̦͕t̹͖͔̖͘r͚̠̰͍͚̹ụ̸̭͍͕̯̹̙s̩͓̼̲̲͉̹t̰.̴͈̖͙̜̲

We will never forget the "Heavenly hundred"
Failures lead to success.
Post

Re: Very Superstitious

#4
Shuul wrote:
i would like to see "suboptimal" behaviour emerge out of the AI,
I dont belive that AI in LT will be able to create "optimal" plan anyway. It is simply too complex for it to handle on perfect level, there are too many factors and many different NPCs that will change those factors with their actions, so no need to worry i think.
And its just the way how it should be.
yeah, but that suboptimal paths are out of incomplete information and not out of "quirks" of the AI
Post

Re: Very Superstitious

#5
Cornflakes_91 wrote:
Shuul wrote:
i would like to see "suboptimal" behaviour emerge out of the AI,
I dont belive that AI in LT will be able to create "optimal" plan anyway. It is simply too complex for it to handle on perfect level, there are too many factors and many different NPCs that will change those factors with their actions, so no need to worry i think.
And its just the way how it should be.
yeah, but that suboptimal paths are out of incomplete information and not out of "quirks" of the AI
Player will be unaware of how "precise" information do NPC possess and due to dynamic world this info will change quite fast i think and that will create discrepensies between "plaaned" value(information) and what NPC will really get. Player will not see the difference most of the time, and if NPC actually gets what he planned then we can just call him smart and a good planner (just like some people IRL)
I̲̩̳̺̩̫n̵̻̘͚͖̗͎ͅ ͢J̜̬̗̦o̩̘̦̪͕͉ͅs͞h̞͘ ̯̹͈͙w̯̙̥e̱͉ ̬̙̘̭̯̦͕t̹͖͔̖͘r͚̠̰͍͚̹ụ̸̭͍͕̯̹̙s̩͓̼̲̲͉̹t̰.̴͈̖͙̜̲

We will never forget the "Heavenly hundred"
Failures lead to success.
Post

Re: Very Superstitious

#6
It all depends on the heuristics of the system that has been created. If Josh has a well modeled cost->action->state system then an AI can be modeled to use less accurate heuristics in its planning algorithm to produce a less than 'optimal' plan. Obviously, more accurate heuristics will produce more 'optimal' plans. Of course, there are a miriad of ways in which this could have been implemented which leaves us with nothing but questions. How many AI entities can be planned for at once? How far into the future can the AI plan? Over how large an area can an AI plan for? Can an AI expand it's goals? (become more ambitious). We'll just have to wait and see what this game can do.

Overall, AI is a fascinating subject.
Post

Re: Very Superstitious

#7
Agreed, immotus, and I have all those questions as well (plus a few more :)).

That's not a criticism -- just the opposite. The capabilities Josh has mentioned so far are very exciting because of the gameplay and just general coolness possibilities they enable. I can't not wonder how far down that rabbit hole will go!

For now, the main question I had here was what other interested parties think about superstitious behavior emerging from NPCs -- how much is just right?

I've added a poll to the thread to get a sense of reactions to this possibility.
Post

Re: Very Superstitious

#9
What kind of superstitions are likely to arise?

An AI spinning when they undock is unlikely unless it's part of a sensor sweep in search of rich asteroids, bounties or the likes.

Repeatedly going to the same asteroid field because an AI believes it's more likely to generate rich rewards, rather than scanning for rich fields all the time is more likely. Unfortunately, it's also a lot less noticeable to the player.

Posting contracts in a certain station because of high traffic or a high density of AI interested in completing contracts for rewards may also be superstition, but also not really noticeable to a player.

What would cause an AI to perform an action that is completely unrelated to their plan in the first place?

Perhaps the AI could be coded to 'take a break' from their plan every now and then. During this time, the AI could perform an action such as making a pilgrimage to a popular system, or go to sleep in a station or random walk around their local system. During these actions that don't relate to their plan, some event may occur that aids or obstructs their progress, allowing the rise of superstitious behaviour.

To help make superstitious behaviour noticeable to the player, the AI may need to share tips to other AI about how they have made their plan more efficient. This could be things like
"I found the combination of this mining tool and that scanning suite turns up rich ore fields more often" or
"I've found that becoming inactive while the player is nearby causes me to get shot all the time"
Or in crazy situations:
"I visited Star System Paradise for a few days and when I came back, asteroid fields were more lucrative"
Basically, "Unrelated action a + unrelated action b +... unrelated action n = detrimental effect x or positive effect y"
Post

Re: Very Superstitious

#10
Superstition is, after all, just incorrect pattern recognition. You can get the online equivalent of a PhD at wikipedia's pattern recognition page, but what it comes down to is basically:

How likely are NPCs to associate N seemingly unrelated events and act upon the association?

Also have a read in the reinforcement article, specifically about reinforcement schedules. Will NPCs respond naturally to different reinforcement schedules (e.g. how humans respond so much more strongly to variable reinforcement as opposed to constant reinforcement). To be exact we're dealing with operant conditioning here (where action by actor > response from environment) rather than classical conditioning (unrelated/neutral event 1 > positive or detrimental event 2)

If LT could simulate operant (as well as classical) conditioning, I'd recommend it for psychology classes. Lol.

As far as technicals: I don't think LT can simulate (and remember) something as subtle as veering in a certain direction or rotating the ship by that many degrees (too many factors to consider), but realistic examples I can come up with:

- Mining ship alone in belt - gets shot multiple times. Mining ship with an escort - gets shot less = NPC always brings escort to mining belt, even in a relatively safe system. ==> "correct" pattern recognition
- Trading with faction X - finds better ores during mining session after = NPC always prefers to mine after having done deals with faction X ==> "incorrect" pattern recognition

Some textual templates (for comms, etc):

I always (do action) after (having done action) with (person or faction). I tend to be lucky then.
It's been in my experience that (do action) after having (done action) is no good.
I don't like to (do action), but when I do, I never do it in (sector/region). Nearly got me killed once.

Both are valid examples of pattern recognition. It's the human perspective that determines whether something "counts" as superstition.
Post

Re: Very Superstitious

#11
Reinforcement schedules as game design have become a staple topic over at Gamasutra, especially for mobile games where getting money from the player quickly is prioritized. "Games as Skinner boxes" basically has its own endowed chair whenever Games Criticism is taken seriously.

On the example I gave of spinning in a circle and thinking that this attracts well-loaded merchants, it wasn't meant to be perfectly representative of LT reality. It was just a quick example of a simple action and a world-effect that an NPC might think were connected. More LT-realistic examples of possibilities where superstitious behavior might emerge are welcome.
jimhsu wrote:I always (do action) after (having done action) with (person or faction). I tend to be lucky then.
It's been in my experience that (do action) after having (done action) is no good.
I don't like to (do action), but when I do, I never do it in (sector/region). Nearly got me killed once.

Both are valid examples of pattern recognition. It's the human perspective that determines whether something "counts" as superstition.
I respectfully disagree with that. I think there are innumerable examples in the real world where -- unless one is a practicing "everything is connected" Buddhist -- some events are clearly not causally connected. Believing that they are is, I think, properly understood as superstitious behavior. (This isn't an assertion intended to go after any particular beliefs of anyone here, BTW. I'm speaking generally.)

This may be even more true in a constructed reality like that of a computer game, where the set of possible actions is constrained to a finite set by what the programmer allows through code. Emergent effects may be hard to predict, but atomic actions are still limited, as are the specific world-effects with which a character action can be correlated. Because in many cases it can be possible to know, through the logic of the code, what caused some world-effect to be produced, it's possible to know whether the character's action was the primary cause for generating that effect.

If it wasn't, and if the character believes it was, then that is a superstitious belief by definition. As the Pythons might say, " human perspective don't enter into it." ;)

All of which is an interesting (really!) but somewhat abstract argument. Regardless of what one calls it, it sounds like NPCs may be able to form causal beliefs. If so, then they might think some particular action they took caused a particular effect that they observe.

Depending on how widespread that it, it could make the characters populating an LT universe capable of sub-optimal behavior.

And that is very interesting indeed.
Attachments
Action-Result.png
Action-Result.png (154.1 KiB) Viewed 2596 times
Post

Re: Very Superstitious

#15
I'm cautiously in favour of this, but replicating human-like irrational behaviour is a pretty complex task. I could easily see this experiment being removed because ai superstitions are just too much like buggy ai.

I mean technically you might think its just pruning ai failure modes to work in interesting ways, but superstition is more than that.

You really need a good mechanism for transmitting beliefs.

Also bear in mind that even superstitious humans who rely too much on luck will fall down if they don't have the capacity to check their results against the real world. You don't want to make NPCs *stupid*.

Online Now

Users browsing this forum: Google [Bot] and 2 guests

cron