[M] monster evolution through genetics

[Kendil]

I guess they could evolve a sence for when to use which skill. So they just not skillspam, but take into enemy position and energy etc into acount. They'll have to have some sort of that anyways, as they'll get new builds, and you can't spend all your time doing costummade AI. Especially if you don't really know what build the AI will have.
What I'm trying to say is that the AI must know how to use their builds. Which is kinda obvious... As I really have no idea of how programing works, there's a big chance that I'm rambling...

Anyways, I was happy when I saw the title again. It really is the best idea ever.

[Prince Zorkian]

This is a really awesome thread. I think the only detail that needs to be kept in mind if they do this is selective area control. Right now, we now that we encounter river Skales in Pre-searing who were all Elementatist and we encountered Bog Skales who were all Necromancers in Kryta, so build wise they were all very similar. In Prophecies, those creatures only dropped a certain collectible item (not including a random weapon or gold), but when we encountered Skales again in NF, at least in Istan, those Skales (not including the Kuskales in Kourna) had up to three options to drop for collectible, but you encountered multiple classes (each different skale name was a different class with a possibility of 4 different classes in one group). For this idea to really work (and I don't have any programming knowledge, so I may be completely wrong) I would think you would just need to randomize the builds similar to how they randomized the drops. A Skale NEVER dropped anything other then those three collectibles (once again, excluding money and other random "all creatures" may drop these). But even within that group, Istan creatures NEVER dropped Vabbi Keys (at least as far as I know). So we know that the developers can limit what a specific group can do, so why not give all Elementatist Skales outside Ascalon 5 or 6 skills that they can choose from, but only have 3 total skills available. Maybe make one or two skills (depending on area) absolute. All Elementatist Skales above lv 4 have Flare, but only a few have meteor Shower or Ice Spear and when we get to Kryta, all Elementatist Skales have Flare and Lava Arrows, then the rest are random. For those worried about Farmers, they are farming for certain items. It is not that hard to adjust your skill set to take into account for all of Fire magic rather then just certain exact skills.

[TedTheShred]

Originally Posted by Kendil

I guess they could evolve a sence for when to use which skill. So they just not skillspam, but take into enemy position and energy etc into acount. They'll have to have some sort of that anyways, as they'll get new builds, and you can't spend all your time doing costummade AI. Especially if you don't really know what build the AI will have.
What I'm trying to say is that the AI must know how to use their builds. Which is kinda obvious... As I really have no idea of how programing works, there's a big chance that I'm rambling...

Anyways, I was happy when I saw the title again. It really is the best idea ever.

Great suggestion.

Early monsters would skill-spam randomly, though the selective pressure would naturally build techniques. For example: monsters that use Gash after sever artery would have a higher fitness function than those who reversed the order. Higher fitness function would increase the likelyhood of selection.

The same would apply to target selection. Monsters targeting spellcasters and monks would develop higher fitness functions than those targeting warriors. Or perhaps completely unique and effective behaviors would emerge.

Different behaviors for different professions would also likely emerge. For instance: dervishes that position themselves next to multiple enemies would output more damage and their fitness function would increase, they would be selected more often, and the behavior would proliferate.

Flock activities would also emerge. Collectively targeting a unit would increase the fitness function of members of a single gene pool. Coordinated body-blocking to prevent players from kiting would increase damage output, but may reduce staying-power.

The beauty of all these systems is that they are easily regulateable and highly adaptive. If a new behavior is needed it will arise naturally; developers won't need to go in and add the behavior. Adjusting fitness functions will adjust the effectiveness of behaviors and the likelihood that they will emerge. Developing this singular system of genetic evolution might be a cumbersome task but it will negate the need to ever again balance monsters and protect against exploits, that evolution will do it for them.

Oh and, agreed. Best suggestion ever.

[Yawg]

I've been dreaming of changing monster builds ever since first solofarms were discovered in early days of GW1. But haven't thought about genetic algorithms... this is PURE AWESOMESAUCE IDEA!

The less repetition and more surprises in my game the better. I was always a full supporter of increasing randomization elements in PvE games to increase replayability and to keep the game challenging, but this idea... is AEONS BETTER than just randomization!

Now only thinking if it's not too complicated if mob groups are highly diversified/balanced as opposed to Prophecies style 1-2 kinds of mobs per group. It's easy to imagine how it would perform for simple mobs like river skales or cave trolls, but what about mobs using actual teambuilds? Maybe it shouldn't matter as the GW2 gameplay in explorable areas will be more like traditional MMOs not like GW1, more fights with smaller groups or 1v1'ing than full party vs large groups of mobs.

[Shadow Ritualist]

What if they recorded their actions and what they did in a battle. Then parts would be randomly selected to continue, such as the example of using a skill after another it works with, so then it carries those skills over and what it did with them, the order it used them in.

So if it used hex removal after having gotten a hex on it, it will remember this behaviour, assuming it's fitness is better than one that didn't.

[HawkofStorms]

Maybe keep this for only some creatures.

Like Fleshweavers... then it would make sense lore wise... sorta

[TedTheShred]

Originally Posted by Shadow Ritualist

What if they recorded their actions and what they did in a battle. Then parts would be randomly selected to continue, such as the example of using a skill after another it works with, so then it carries those skills over and what it did with them, the order it used them in.

So if it used hex removal after having gotten a hex on it, it will remember this behaviour, assuming it's fitness is better than one that didn't.

This is a slight misunderstanding of how this would work. On the right track, but still a misunderstanding.

Think of a string of rules (or genes): One gene determines the conditions necessary to cast a particular skill. Once these conditions are met, the skill is activated. An example set of these conditions would be: I am targeting myself, I am stationary, my heath is between 50 and 100, my energy is between 10 and 30, I am being attacked, I am bleeding, I am hexed. Once all of these conditions have been met, and given that the skill is not recharging, the skill will be cast. These conditions are like switches and dials, when set to these positions will either positively or negatively affect the fitness function of the monster.

Monsters with higher fitness functions are more likely to survive and pass on these strong genes. Subsequent generations will mutate this gene, as well as others, and these mutations will either increase or decrease the fitness function of the monster.

There is no need to 'record' previous battles. Genes that tend to increase the fitness function of a monster will be passed on, and the behavior will spread.

Originally Posted by HawkofStorms

Maybe keep this for only some creatures.

Like Fleshweavers... then it would make sense lore wise... sorta

It makes logical sense for all creatures. This is a similar (if accelerated) version of what happens with animals in the real world. Additionally, it makes sense that with the litany of skills in Tyria, that monsters and enemy races would adapt their builds to best defend against the players.

[HawkofStorms]

Originally Posted by TedTheShred

It makes logical sense for all creatures. This is a similar (if accelerated) version of what happens with animals in the real world. Additionally, it makes sense that with the litany of skills in Tyria, that monsters and enemy races would adapt their builds to best defend against the players.

The bolded bit is what bothers me.

Bad "fake geneticist technobabble" always kinda annoys me. So... makes sense for creatures that are designed to be adaptive (like Fleshweavers).


Although, lore and logic should always take a back seat to good game design ("WTF you should be dead, I just hit you in the head with a sledgehammer", would be a terrible game). So, yeah, this is still totally a good idea for the entire game.

[Gregor Lightbringer]

Originally Posted by HawkofStorms

The bolded bit is what bothers me.

Bad "fake geneticist technobabble" always kinda annoys me. So... makes sense for creatures that are designed to be adaptive (like Fleshweavers).


Although, lore and logic should always take a back seat to good game design ("WTF you should be dead, I just hit you in the head with a sledgehammer", would be a terrible game). So, yeah, this is still totally a good idea for the entire game.

Yeah. Just imagine hitting a slime monster like that with a hammer. Instead of it going SPLAT! it splats all over you only to envelop and eat you. Thats evolution for ya.

[TedTheShred]

Originally Posted by HawkofStorms

The bolded bit is what bothers me.

Bad "fake geneticist technobabble" always kinda annoys me. So... makes sense for creatures that are designed to be adaptive (like Fleshweavers).

Well, when I mentioned that it was accelerated, I was simply noting that while the process is essentially the same, the nature of being a videogame accelerates the process.

In real life, the time between generations can be years. In the game, it could be minutes. The degree of change in the environment is also much more dramatic. Effective builds spread very quickly through PvX, and what was once a working strategy is now pitifully ineffective.

Those two conditions simply create an accelerated version of the same process when compared to our personal perception of the process in real life.

Also, note that I mentioned that the monsters in Tyria are probably all at least kind of smart. For instance, if a specific skill becomes wildly ineffective against someone they'll stop using it and start using something else. Like if a monster sees another monster use Flurry against a Necromancer using Spiteful Spirit; after the carnage, that monster would probably be hesitant to use Flurry in the presence of Necromancers. Genetic evolution would mimic this normal and plausible behavioral shift.

[Craywulf]

I'm all for an evolutionary A.I.

This would give the game so much more re-playability and encourage player grouping.

I'm not the least bit concern about beginner area adapting either because I think it will breed better new players and bring them a little closer to the PvP game than it currently does now. There isn't a single AI that prepares a player for PvP in terms of learning the metagame.

I hope ArenaNet considers this idea, as I think it would make the game incredibly more competitive. Which is something ArenaNet prides Guild Wars to be.

[Corpsesarefun]

While its probably a little late in development for this to be implemented in GW2 you inspired me to create an evolution system (as a substitute for a monster levelling system) for a turnbased RPG im currently working on as a hobby

Rather than your system it works by applying a random change to the stats or attacks of each monster when it spawns, then the game checks to see if this change is a good thing and if it is the chance of that change reoccuring in future spawnings goes up by 10%.

[raspberry jam]

Originally Posted by dakka dakka

I like your idea, is very original and looks good on paper, though it may look like a code nightmare to programmers

Nope, the algorithm itself is rather easy actually. Getting all the parameters right (such as the change of fitness when the new build/behaviour performs well or fails to do that) is a whole different thing though, but that's a math nightmare, not a programming one.

One flaw in the suggestion is that it fails to consider evolutionary dead-ends (situations with low fitness where every permissible change means even lower fitness) due to topological granularity of the skills (yeah yeah, I know ). Hopefully those would happen so close to the target as possible but if they don't, you'd end up with monsters that are either completely retarded or retardedly overpowered. In either case you'd need to reset those monsters.

Other than that... This suggestion is the best I've seen on any board like this. I think I'll steal it.

[draxynnic]

Originally Posted by Kendil

But, couldn't this be overcome by farmers? If you use a gimmick build to defeat (for example) scales. Then they evolve a counter to that. What if the farmers just creates a second gimmick build, and that builds counter is the first gimmick build strong against.

Is it a too small a chance? If not, is it a way to counter that?

I'd call that not an issue. ANet doesn't consider farmers to be a problem, they just don't want it to be too easy.

Raspberry, I suspect your problem is dealt with by tmakinen's suggestion that the least "fit" build be discarded completely - those evolutionary dead ends will go extinct.

On the issue about monster variability (especially at the high end where they'd be pushing for maximum effectiveness): I suspect this is where the constraints come in. ANet will want to make sure that certain monsters have certain flavours for lore reasons, and this could be achieved by fixing certain skills. Consider drakes, for instance - the drakes on the Fire Island chain are obviously designed for running in and blowing up their foes with close-range fire spells, so their constraints should be set so that they never end up with a long-range blasting build even if such a build is considered "fitter". If the system is anything like GW1's, even just constraining the elite should make a big difference - even with no other constraints, a monster which is fixed with Mind Blast as its elite will probably come out with a quite different set of supporting non-elite skills to one with Savannah Heat, which will be different to one with Double Dragon, which will be different to one with Searing Flames.

On the whole, looks like a good idea, if it can be implemented in a smooth manner.

[raspberry jam]

Originally Posted by draxynnic

Raspberry, I suspect your problem is dealt with by tmakinen's suggestion that the least "fit" build be discarded completely - those evolutionary dead ends will go extinct.

Um... no. An evolutionary dead end is where you appear to have reached maximum fitness when you really are nowhere near it. In other words, these builds would be the most fit ones in the pool, yet they could be much better (and by "better" I mean better for their purpose, in other words, not over- nor underpowered).

This problem could be caused by bad rules for random mutations. Say for example (very simplified) that the rule would be to change one skill. But any one-skill change to "build A" would produce a worse (less fit) build. Thus build A would never change because it would out-compete any build close enough to it.

This is a common problem with this type of algorithm and you'd need to be very careful to avoid them completely. It's a minor issue though, as long as ANet would watch out for this and reset the skills if it gets too bad (though that may not even be needed).

[TedTheShred]

The evolutionary dead end is a valid issue, but I think it's something of a red herring.

Take a look at the state of PvE in Guild Wars, or rather the effect of PvX on the state of PvE. Players are rabid in their pursuit of exploits. If a build has a weakness, they will find it, they will use it, and they will pass it on to everyone they can. In essence, the natural state of the player community is one that aims to reduce the fitness function of every and any monster.

If an evolutionary dead end is reached, if progress is so stilted that it's indistinguishable from inert, then players will notice. Those rabid, exploitative players will descend on a build that they can so regularly defeat with such vigor that this monster's fitness function will plummet. At that point a new skill will eventually emerge; one that in most scenarios would be detrimental to the build. However, in the case of build exploitation, where a build is so unfit, that skills running counter to the synergy of the build is a benefit; as that synergy is a failure.

On top of that, encountering an inert build will increase the homogeneity of the exploitative player-builds. That homogeneity will open the door for a unique scenario. Homogeneity decreases immunity. A single moderately successful change in the monster's build will rapidly increase the fitness function, because that small up-tick in effectiveness will affect every single player using that exploitative build. Once that path is open, the system will take it's course: skills that complement the old ineffective build will be selected less often than skills that complement the new effective counter-skill.

**Edit**

I'd like to add a little bit more. If one of these evolutionary dead ends does occur, it's kind of a non-issue. If a build is working for this set of monsters, there is no urgent need for a change. There will remain some variety, albeit ineffective variety, but variety nonetheless. That a monster *could* be more powerful, but that power is unattainable because it can't mutate drastically enough, is not the goal of the system.

That dead-end is simply a niche. There is no reason for drastic change unless that niche is exploitable, and as I illustrated above, when it is exploited the system will react accordingly.

[Shadow Ritualist]

Well, you could probably avoid exploitation by making the amount of change dependent on the fitness. So if people are just destroying that by exploiting it, then it will change quickly and drastically.

[tmakinen]

Hehe, I had completely forgotten this thread already. Concerning evolutionary dead ends, you seem to have it mostly worked out already. I touched the subject very briefly in the OP or thereabouts while speaking about the devs' ability to monitor population health in real time through the fitness function. If a particular population gets stuck to a local maximum and the player base notices this and pounds on it with gusto, the fitness of the entire population plummets and signals to the devs that maybe something should be done.

However, the situation is less severe than one might expect. Simulated annealing solves the local maxima problem by taking random steps around, favorable steps always and unfavorable with a probability of exp(-E/kT) where E is the energy malus of the step, T the system temperature and k a coupling constant. Make T inversely proportional to the volume of the population in parameter space and there you have a self-regulating system (as the entire population gets crammed towards the local maximum its temperature rises, creating an apparent (entropy based) evolutionary pressure that pushes individuals outwards from the maximum). Even that may be over-engineering it since the parameter space is extremely shallow. Gimmick builds are such just because they can be completely trashed with a simple countermeasure, usually just one particular skill. Thus, using the GW1 system as an example, killing a permasin farmer requires just one disenchantment skill (which the algorithm will find very fast), not a unique bar of 8 skills (which would be beyond the reach of the algorithm).

Because I don't feel like starting a new thread, I'll just add that there is an even simpler system for a PvE meta, mirror AI that has been used, e.g., in flight simulators. Basically, the computer observes the player(s) and feeds their behavioral patterns back at them. In this case it would mean that monsters gradually 'learn' skill bars from the players (skill usage statistics is already internally collected in GW1). The implementation could be as simple as every now and then swapping a random skill on a monster to whatever happens to be the FotM among the players. However, this system would be inherently easier to abuse if the players found out how it works.

[raspberry jam]

@TedTheShred Remember that fitness isn't the strength of the build, but the suitability for the level of challenge that the area is meant to present. Overpowered enemies in an area which is supposed to be easy is just as wrong as enemies that are so easy that they become farming targets. And actually I expect monsters that are too strong will be harder to detect (and less likely to be fixed) than those who are too weak.

Originally Posted by tmakinen

Simulated annealing solves the local maxima problem by taking random steps around, favorable steps always and unfavorable with a probability of (...)

Then I have no objections anymore. It's perfect and I wish they'd use it.

Originally Posted by tmakinen

Because I don't feel like starting a new thread, I'll just add that there is an even simpler system for a PvE meta, mirror AI that has been used, e.g., in flight simulators. Basically, the computer observes the player(s) and feeds their behavioral patterns back at them. In this case it would mean that monsters gradually 'learn' skill bars from the players (skill usage statistics is already internally collected in GW1). The implementation could be as simple as every now and then swapping a random skill on a monster to whatever happens to be the FotM among the players. However, this system would be inherently easier to abuse if the players found out how it works.

Easier but still pretty hard to abuse, because a FotM is a FotM since it works... If players wanted to abuse it, thousands of them... of us would have to agree to use the same rather bad build in the hope that monsters will start using it so that we can farm them. Sounds unlikely.

[Kendil]

Originally Posted by tmakinen

However, the situation is less severe than one might expect. Simulated annealing solves the local maxima problem by taking random steps around, favorable steps always and unfavorable with a probability of exp(-E/kT) where E is the energy malus of the step, T the system temperature and k a coupling constant. Make T inversely proportional to the volume of the population in parameter space and there you have a self-regulating system (as the entire population gets crammed towards the local maximum its temperature rises, creating an apparent (entropy based) evolutionary pressure that pushes individuals outwards from the maximum). Even that may be over-engineering it since the parameter space is extremely shallow.

Could someone explain that to me again?

As I didn't really get what tmakinen said, I'll say why I think the dead ends won't occur. New builds are created at all times, you don't just change one skill, but a whole new build could be randomized, or two builds could be merged to one.
As a new build could be totally random (with the restriction that's set for the specific population) it would be impossibe to come to a dead end, as a new build could be drawn from scratch, and if it's fitter the new build would be the one to spread. Or at least a mutation of the new one.