|
ToxicSlurpee posted:Endless chaff was actually one of the things I've been trying to avoid. ADOM in particular is the one I think was the worst for this, partly because of how the chaff scales. For those that haven't played it some things in the game can summon a horde of monsters every turn for several turns in a row and I do mean a horde. Jackalwares will tend to immediately summon like 30 of the drat things immediately after being stabbed once. Jackals individually aren't tough at all and by the time you're like level 10 they can't even hurt you. This and the ridiculous amounts of equipment destruction are what made me stop playing ADOM. I still enjoy the game, but I can't play it because it's just absurd.
|
#
?
Mar 31, 2015 16:57
|
|
|
|
| # ? Apr 25, 2024 06:50 |
|
|
Edwhirl posted:This and the ridiculous amounts of equipment destruction are what made me stop playing ADOM. play beastmaster or monk and have fun instead.
|
|
#
?
Mar 31, 2015 18:33
|
|
|
ToxicSlurpee posted:Endless chaff was actually one of the things I've been trying to avoid. ADOM in particular is the one I think was the worst for this, partly because of how the chaff scales. For those that haven't played it some things in the game can summon a horde of monsters every turn for several turns in a row and I do mean a horde. Jackalwares will tend to immediately summon like 30 of the drat things immediately after being stabbed once. Jackals individually aren't tough at all and by the time you're like level 10 they can't even hurt you. I don't think the base scaling is nearly as severe as you mention. I think the jackals were a special case due to a bug that made the modifier x10 or x100 for their scaling off kills or something like that. Also I don't really think most roguelike AI needs to be improved much (or at all). I know ADOM and Crawl (at the minimum) have certain enemies that do run when injured or go berserk (and chasing runners is one of the more annoying combat situations in either game imo). Also how is AI supposed to determine how strong you are? Just hp? I mean anything beyond that is going to end up with odd situations where people try gaming the system by manipulating whatever numbers the AI checks (even with hp that could be possible to an extent, depending on where your breakpoints are). Maybe for something like set encounters, but when there's 20 dudes on the screen I don't really want to have to worry about their behavior patterns, on top of having to know their attacks/defenses/abilities all while calculating mine and also worrying about managing resources and navigating the map. I guess it would have to be something I saw in play to see if there's a big improvement gained, but AI would be near the bottom of my general roguelike complains.
|
|
#
?
Mar 31, 2015 20:01
|
|
|
I really like how Crawl uses popcorn enemies. It's the same design philosophy as Doom's bestiary in that some enemies serve as nuisances and aim to distract you from far away more than anything else (zombies), while others are slow to move and attack but can significantly block your escape routes (cacodemons & demons), and others aren't necessarily highly damaging but they do a great job in limiting which weapons you can use (lost souls and rocket launcher). It all works beautifully. Crawl's popcorn enemies often serve their purpose sufficiently just by blocking your escape route. There's also good synergies with higher level enemies that can cause bloodlust. Tome is just booooring.
|
|
#
?
Mar 31, 2015 20:15
|
|
|
alansmithee posted:I don't think the base scaling is nearly as severe as you mention. I think the jackals were a special case due to a bug that made the modifier x10 or x100 for their scaling off kills or something like that. Have you ever played Sil? It has interesting basic AI that doesn't bog down the game.
|
|
#
?
Mar 31, 2015 20:28
|
|
|
alansmithee posted:I don't think the base scaling is nearly as severe as you mention. I think the jackals were a special case due to a bug that made the modifier x10 or x100 for their scaling off kills or something like that. I think Biskup is working on the scaling but yeah super jackals were part of a bug but even so I remember a few particular runs where I ran into absurd numbers of dark elves that just summoned spiders everywhere. The Big Room in particular ended up just filled with spiders at one point. I had killed I think at least 900 of each type at one point and the things just wouldn't die by that point. It doesn't happen every game but it can, especially if you find yourself running into spiders and dark elves. The AI responds to what it sees the player do and how it interacts with the player actually. The main thing right now is damage; whenever the player hits a creature everything that can see it will take note of how badly the player hurt the other creature. The damage is compared to how many hit points each creature knows it has. If the creature notices that the player is capable of hurting it very badly in one hit, especially if the player can just plain kill it in one hit, the creature will be more cautious around the player or just avoid him entirely. These numbers are based on how much the creatures see the player doing. Creatures don't just magically know every capability the player has. They can see how hurt the player is but everything else they need to figure out by seeing it. The idea of using neural networks is to let the creatures learn some things along the way and respond appropriately. A freshly spawned creature shouldn't just automatically know how powerful the player is if they haven't seen him do anything. Other creatures just completely ignore this; think mindless things or things that are fearless. Creatures also respond differently if they're desperate for some reason or another. You're right that roguelikes can work with very simple AI. The biggest goal really is to make the dungeons feel more alive and the creatures less binary. I'm a sucker for little details.
|
|
#
?
Mar 31, 2015 20:53
|
|
|
ToxicSlurpee posted:If the creature notices that the player is capable of hurting it very badly in one hit, especially if the player can just plain kill it in one hit, the creature will be more cautious around the player or just avoid him entirely.
|
|
#
?
Mar 31, 2015 20:57
|
|
|
If it's not outside the scope of the project, you could do swarm AI as well, where some monsters will do threat assessment for themselves, while some do it for their pack and will happily attack because they believe the swarm will kill you even if they personally will likely die.
|
|
#
?
Mar 31, 2015 20:58
|
|
|
Awesome! posted:chasing around trash mobs that i can 1shot but that wont engage me sounds absolutely awful I'm considering how to deal with that actually. I know the mentality in a roguelike is "kill absolutely everything" and I'm debating on to just how far I should take these things. Right now an uninjured creature that knows it can be one-shot by the player is cautious but doesn't run away. Granted that can be avoided by just not spawning piles of trash mobs but that's a design thing I haven't finalized quite yet. Jeffrey of YOSPOS posted:If it's not outside the scope of the project, you could do swarm AI as well, where some monsters will do threat assessment for themselves, while some do it for their pack and will happily attack because they believe the swarm will kill you even if they personally will likely die. Currently swarm AI is just stuff that is set to be suicidally reckless and spawned in groups. That isn't actually beyond the scope of the project and I've been thinking about how to handle it but that relies on stuff that hasn't been fleshed out yet. I tested things that have their numbers set to fearless and permanently hostile and they swarm the player just fine. Actually that gets into the gubbinz in the AI program. Some flags can be set to make things that do the standard "run at player -> try to kill player" and the weighting in the network can be set to make the creature more or less cautious and more or less brave. Not everything will consider running away while some things are far more skittish. I actually also liked the idea of things that just flat out won't fight the player and try to run away. Think stuff that you hunt; prey animals and the like. ToxicSlurpee fucked around with this message at 21:33 on Mar 31, 2015 |
|
#
?
Mar 31, 2015 21:31
|
|
|
Things that run away should actually run away. Not return in 20 minutes. Like run for one of the two stair cases and then disappear entirely from the game world.
|
|
#
?
Mar 31, 2015 21:35
|
|
|
ToxicSlurpee posted:The problem is that every time you kill a jackal it makes all future jackals more powerful. Most monsters work like this but there are very common summoning monsters that summon a gently caress load of one specific type of monster. You can end up with stupid bullshit like spiders with 900 hit points or jackals that can no longer miss and God help you if you run into something that summons 30 of something that also summons monsters. The challenge then is dealing with these massive hordes of monsters and how tough they can become but for the most part they just become damage sponges. It's incredibly tedious especially considering that the AI is of the "run at player -> attack player" variety for the most part. It gets even worse if you can't carve through the gigantic blob of sponges to get to the drat thing summoning them before it just summons more. All this used to be true but it's not anymore. Uberjackals no longer exist and summoners have been made less annoying with their summons.
|
|
#
?
Mar 31, 2015 21:39
|
|
|
CountingWizard posted:Things that run away should actually run away. Not return in 20 minutes. Like run for one of the two stair cases and then disappear entirely from the game world. I'm planning on that actually. I'm working on a goal system and better pathfinding for that sort of thing. My plan is also for the creatures to still give at least partial XP if the player beats the poo poo out of it but it gets away. It's on the board I just have other stuff to get working before that. That's also where learning comes in; creatures don't magically spawn knowing where the retreat routes are and explore to find that sort of thing. Right now semi-random exploration and searching for food sources is in but complex enough path finding to do that sort of thing is not.
|
|
#
?
Mar 31, 2015 21:57
|
|
|
ToxicSlurpee, I would also be interested in a preview of that paper. I'm particularly interested in the specifics of the neural-network stuff you're using, i.e. the features you use, the network topology, etc. I would not generally expect a neural-network based AI to be all that successful for a roguelike game, but it seems that you've gotten some decent results so far. I have a tendency to start writing roguelike engines and then getting bored before making any content for it. The most complex AI I've implemented used a goal-oriented action planning (GOAP) approach and it worked out pretty well, although I could see that as the complexity of the environment and the possible actions increased, you would need to build a pretty complex ontology/action schema to get good AI results. Neural network approaches have the benefit of not necessarily needing everything spelled out in explicit ontologies, but they also seem harder to analyze and tweak, and without some protections they could be as susceptible to thrashing (i.e. a monster switching back and forth between fleeing and attacking states) as other approaches.
|
|
#
?
Apr 1, 2015 02:22
|
|
|
Mr. Wednesday posted:ToxicSlurpee, I would also be interested in a preview of that paper. I'm particularly interested in the specifics of the neural-network stuff you're using, i.e. the features you use, the network topology, etc. I would not generally expect a neural-network based AI to be all that successful for a roguelike game, but it seems that you've gotten some decent results so far. Thirded! It's been a while since I've been hip deep in robust AI discussion, I kinda miss it. I think it rules when someone comes into this thread and drops mad  about Roguelikes, it's completely the best, it's why I love this thread (and its ancestor) and I hope you end up making something super awesome and playable. Game AI is such a crazy and wild flock of multi-hued birds. Smart AI and the illusion of Smart AI are two entirely different things to us creators but to the players they are often the same, or worse the players prefer the fake smart AI rather than the genuinely smart one which plays with a methodical, efficient yet unengaging style. When I think about genuinely smart-feeling bad guys, I always think about the wizard fights in Baldur's Gate. Motherfuckers always had a triple-stuffed bag of tricks with contingencies, clones, fall backs, potions and other sorts of clever planning crapola that made them feel infuriatingly smart and tough. The truth of it though was that the heavy lifting was done by a smart human coming up with clever scripts, and the AI itself is simply responding to game states without any real emergent behavior. Dungeonmans lets you carry some degree of player state from game to game, and in a small way you change enemy situations by having your killer get promoted to dungeon boss. But could it work the same for enemy AI? What if the Scrobold Warren was a stinky hive of learning, and it evolved based on player input? We could go on for pages about the possibilities, but to the average end user it would feel like "Man my favorite build sucks now because all the AI is too smart for it, so I have to play a different class a few times to make them stupid again so I can go back to playing what I really like. This game sucks." 
|
|
#
?
Apr 1, 2015 02:42
|
|
|
madjackmcmad posted:Dungeonmans lets you carry some degree of player state from game to game, and in a small way you change enemy situations by having your killer get promoted to dungeon boss. But could it work the same for enemy AI? What if the Scrobold Warren was a stinky hive of learning, and it evolved based on player input? We could go on for pages about the possibilities, but to the average end user it would feel like "Man my favorite build sucks now because all the AI is too smart for it, so I have to play a different class a few times to make them stupid again so I can go back to playing what I really like. This game sucks." Something I've occasionally vaguely wondered about would be making a "multiplayer" singleplayer RPG where the enemy AI was based on previous players' responses to situations. Basically you bootstrap the game with a basic handwritten AI, then you examine how the player responds to that AI -- when the player is in a situation that looks like X, they perform action Y. And you abstract these responses somehow, and use them to build the next tier of AI. It's "multiplayer" in that many players' actions are being used to train the AI (as you play the game is collecting information on your playstyle). Each time you learn to consistently beat Tier N of the AI, it ranks up to Tier N+1, which was trained based on how players worldwide beat Tier N. I can't imagine the type of game that would allow for that kind of depth of strategy, let alone the logic needed to adapt player actions in one situation to AI actions in a similar but subtly different situation. That, plus my total lack of free time, are why I haven't been pursuing this concept more actively, because otherwise it sounds seriously neat, to me anyway.
|
|
#
?
Apr 1, 2015 02:53
|
|
|
madjackmcmad posted:Game AI is such a crazy and wild flock of multi-hued birds. Smart AI and the illusion of Smart AI are two entirely different things to us creators but to the players they are often the same, or worse the players prefer the fake smart AI rather than the genuinely smart one which plays with a methodical, efficient yet unengaging style. There was a study I read, though annoyingly I can't find it at the moment, where players were asked to rate the AI in a FPS game. One AI was rated much higher than the other - the difference between them was that the agent of the highly-rated AI had significantly more HP than the other one. There was no difference in the actual behavior, but the perceived difference in intelligence was quite large. This is a great example of the conflicting goals that often come up between game designers and game players. Game players think they want to play against 'intelligent' opponents, but in reality they often don't, check out Smart Kobolds for a case study in this, as others have pointed out. What they really want is the illusion of intelligent AI. I'd go on about the implications for game developers but I'm a bit too drunk to pull that off at the moment
|
|
#
?
Apr 1, 2015 02:53
|
|
|
Barring stuff like chess where the AI is sitting in for an actual human opponent, the purpose of a computer game isn't to outsmart the player or even fight by the same rules; it's to create a series of difficult decisions that all test the same skill or tightly clustered group of skills. If you let the computer start really thinking for itself the decision becomes more like a guess, and therefore just as infuriating and unrewarding as actual randomness. e: Instead of one enemy who learns and adapts, consider making each type of enemy behave differently, but as consistently as possible. Then combine them in groups where their respective tactics create unique difficulties for the player (like tank/tarpit enemies that get up close and prevent you from moving alongside ranged kiters, for a simple example.) You can mix up the environment, too -- the point is to create predictable, learnable patterns, and then change them just as the player masters them but before they get bored. Tuxedo Catfish fucked around with this message at 03:03 on Apr 1, 2015 |
|
#
?
Apr 1, 2015 02:54
|
|
|
Tuxedo Catfish posted:e: Instead of one enemy who learns and adapts, consider making each type of enemy behave differently, but as consistently as possible. Then combine them in groups where their respective tactics create unique difficulties for the player (like tank/tarpit enemies that get up close and prevent you from moving alongside ranged kiters, for a simple example.) You can mix up the environment, too -- the point is to create predictable, learnable patterns, and then change them just as the player masters them but before they get bored. This is a good formula for a great game imo. I point at Necrodancer for a prime example of this approach done very right.
|
|
#
?
Apr 1, 2015 03:13
|
|
|
It's also one of the reasons the "boring" parts of ToME aren't a complete waste, although they could certainly still use some calibration: occasionally presenting the player with enemies that aren't a lethal threat alone, but are in conjunction with other enemies, gives them a chance to learn their behavior before they're put in a life-or-death situation. The flip side of that is that for experienced players it quickly becomes unnecessary, but I suspect its effect on accessibility is one of the reasons ToME became so popular so quickly. (Of course, being a roguelike, you could reasonably argue that the player should just HTFU and learn by watching how they die.  )
|
|
#
?
Apr 1, 2015 03:16
|
|
|
Ugh, thread made me want to try ADOM, and the controls are so obtuse. Are there any suggested configurations to make this poo poo easier? More broadly, is this worth the time and effort? I've never even gotten a rune in crawl.
|
|
#
?
Apr 1, 2015 06:27
|
|
|
Lprsti99 posted:Ugh, thread made me want to try ADOM, and the controls are so obtuse. Are there any suggested configurations to make this poo poo easier? More broadly, is this worth the time and effort? I've never even gotten a rune in crawl. ADOM is not a very well designed game.
|
|
#
?
Apr 1, 2015 09:22
|
|
|
Lprsti99 posted:Ugh, thread made me want to try ADOM, and the controls are so obtuse. Are there any suggested configurations to make this poo poo easier? More broadly, is this worth the time and effort? I've never even gotten a rune in crawl. even better in other roguelike games you can choose "adom" as key configuration.
|
|
#
?
Apr 1, 2015 10:18
|
|
|
[CoQ Steam 4-2-2015] *More tiles! *Fixed a bunch of effect overlays not working *Fixed gas colors not being applied properly *Fixed various target pickers not flashing on selected tiles *Fixed a bunch of creatures improperly being flagged as visible hostiles *Fixed a few on-seen triggers that weren't working *Fixed a bunch of broken animations *Fixed paralysed leaving a dangling reference *Removed some excess threading model stuff
|
|
#
?
Apr 3, 2015 02:45
|
|
|
Mr. Wednesday posted:ToxicSlurpee, I would also be interested in a preview of that paper. I'm particularly interested in the specifics of the neural-network stuff you're using, i.e. the features you use, the network topology, etc. I would not generally expect a neural-network based AI to be all that successful for a roguelike game, but it seems that you've gotten some decent results so far. I won't be able to have an overall preview of the paper sadly but I'll be sure to post the paper itself later on when it's more done. It's pretty bad at the moment and I'm still working on some things to prove what I want to do can be done. I need to have a speech together for school about it before the paper so I'm focusing more heavily on that, really. As for the basic overall topology the network actually turned out to be very simple when compared to stuff like handwriting recognition. The short of it is that there is a stack of float values for each tile of a dungeon level (as it stands I don't have creatures moving between levels so tracking more than one of them is irrelevant) that rates the desirability of each tile. This includes things like whether there are things the creature desires there, where the player is and how it feels about the player, and if the tile has a known trap or not. That sort of thing. Tiles that lead the creature toward some goal are also rated more highly. Where the fuzziness comes in is that the creatures look at nearby tiles that can get them to wherever they want to go and pick which one is the best. Creature personalities and the weighting can also affect it. An injured creature is going to put a much larger negative effect on a trapped, dangerous tile than one that is unhurt. Anyway, some of the layers are just pure ratings; tiles the creature doesn't like for some reason are negative while places it likes are positive. Different layers track different things and paths can traverse a particular layer without affecting others. Hence a creature with multiple goals can actually consider tiles that will get them closer to multiple ones. Weights can also be applied to give goals priorities (mind you goals are still being fleshed out because the pathfinding is basic as I type this). These tiles are used for various weighting in the actual processing of creature actions. Deciding just what to do is based on what amounts to single neurons; at the absolute most the decision is based on two layers. More was just unnecessary. It checks the totals and thresholds in a particular order (some actions have a higher priority than others in this way) and if a neuron goes off no other neurons are checked. It actually took relatively few actual neuronal checks and connections to get the creatures going after stuff they could see. Creatures can also be programmed to ignore neurons based on weights. A weight of 0 on cowardice for example makes the creature totally incapable deciding to run from something. The main thing this does is lets the creatures create fuzzy paths through the dungeon. I won't go into the technical side of things as I'm still figuring out how to explain it but other things that affect it are stuff like how good the creature's attacks are. If the creature has a powerful ranged attack but sucks in close combat it will value tiles that allow it to stay away from threats more than those that are close to threats. Personality is another big thing; the way neural networks function allows for fiddling with the weights to make creatures respond more powerfully to things. A cowardly creature is going to rate a tile far more negatively than a brave one. I was more concerned with the creatures having personalities and adapting to their surroundings (as far as they know, anyway) than anything else. But yeah sorry but I won't have a preview available. The paper itself will go up eventually. I think about half of the creatures not being hostile by default would seem about right; this way it seems more like a living, breathing dungeon that the player can make choices in would be interesting. That way a player can do the traditional "murder literally everything" style of play or let stuff live. Granted that also means that experience and equipment must have alternate sources but that I need to tackle yet. Tuxedo Catfish posted:Barring stuff like chess where the AI is sitting in for an actual human opponent, the purpose of a computer game isn't to outsmart the player or even fight by the same rules; it's to create a series of difficult decisions that all test the same skill or tightly clustered group of skills. If you let the computer start really thinking for itself the decision becomes more like a guess, and therefore just as infuriating and unrewarding as actual randomness. Yeah the other side of it is that a roguelike by its nature tends to lead to hordes of monsters going after the player while the player massacres thousands of them. I actually took a far simpler approach to the AI then I had planned on but you know, that's how game making works. The neurons are still there the stuff is just way simpler than the original plan. What I had actually wanted was creatures that more mostly predictable. As in, you know that certain creatures are fearless and want you dead but the details of how they go about that was different. I also wanted creatures that were more predictable than others. I felt like fuzzy logic, hence the approach I took, was the best, especially when you consider that the dungeon can also be dangerous to the creatures in it. I also liked the idea that the player wasn't the only person thrown randomly into the dungeon; I liked the idea of other monsters being thrown in and being just as disoriented and confused as the player. One thing I had actually thought of was creating a sort of hive mind that learned and adapted while the individual creatures were mostly dumb. Does that sound interesting? ToxicSlurpee fucked around with this message at 04:13 on Apr 3, 2015 |
|
#
?
Apr 3, 2015 04:09
|
|
|
Unormal posted:[CoQ Steam 4-2-2015] i bought sproggiwood, how can i get CoQ?
|
|
#
?
Apr 3, 2015 14:12
|
|
|
The Something Awful Forums > Discussion > Games > The roguelikes thread: how can i get CoQ?
|
|
#
?
Apr 3, 2015 15:57
|
|
|
It'll be available on Steam in 2-3 weeks!
|
|
#
?
Apr 3, 2015 15:59
|
|
|
Unormal posted:It'll be available on Steam in 2-3 weeks! 
|
|
#
?
Apr 3, 2015 16:00
|
|
|
inormal when is sproggerwood on ios
|
|
#
?
Apr 3, 2015 19:40
|
|
|
Bogart posted:inormal when is sproggerwood on ios Were trying to launch sproggi mobile and caves of qud and savage mode the same day. Well see how that goes.
|
|
#
?
Apr 3, 2015 19:46
|
|
|
Unormal posted:Were trying to launch sproggi mobile and caves of qud and savage mode the same day. Well see how that goes. Why do you hate yourselves?
|
|
#
?
Apr 3, 2015 20:15
|
|
|
Bobo the Red posted:Why do you hate yourselves? Freehold Games: Maximum Self Loathing
|
|
#
?
Apr 3, 2015 20:21
|
|
|
Unormal posted:Freehold Games: Maximum Self Loathing Is this penance for the seven-twinned lampreys?
|
|
#
?
Apr 3, 2015 20:21
|
|
|
Sir Unimaginative posted:Is this penance for the seven-twinned lampreys? If we don't launch them all simulatenously they all unlaunch and we have to try again.
|
|
#
?
Apr 3, 2015 20:25
|
|
|
Unormal posted:If we don't launch them all simulatenously they all unlaunch and we have to try again. Also your code deletes itself and you have to develop them all again from scratch!
|
|
#
?
Apr 3, 2015 21:56
|
|
|
Unormal posted:If we don't launch them all simulatenously they all unlaunch and we have to try again. 
|
|
#
?
Apr 3, 2015 22:03
|
|
|
Wafflecopper posted:Also your code deletes itself and you have to develop them all again from scratch! Last I heard, Unormal was attempting for an Ultimate release. Still looking for that Amulet of Life Saving...
|
|
#
?
Apr 4, 2015 00:40
|
|
|
Nah, AoLS is for the normal ending. You need a CoQ ring for the Ultimate release.
|
|
#
?
Apr 4, 2015 01:20
|
|
|
Any discounts on the Steam CoQ for those who registered their free version? I'm more than happy to pay whatever the full price is.
|
|
#
?
Apr 4, 2015 01:37
|
|
|
|
| # ? Apr 25, 2024 06:50 |
|
|
Sage Grimm posted:Nah, AoLS is for the normal ending. You need a CoQ ring for the Ultimate release. 
|
|
#
?
Apr 4, 2015 01:47
|
|