WEBVTT 1 00:00:00.080 --> 00:00:02.839 Okay, So let's stick into something that it's way more 2 00:00:02.919 --> 00:00:06.080 fascinating than just making you know, bad guys shoot straight 3 00:00:06.120 --> 00:00:11.199 in games, we're talking about artificial intelligence, but specifically in games, 4 00:00:11.800 --> 00:00:17.079 that kind of hidden intelligence that makes the world's feel alive. 5 00:00:16.960 --> 00:00:19.800 Rights, the stuff making characters seem like they're thinking or 6 00:00:19.839 --> 00:00:21.079 moving believably. 7 00:00:20.879 --> 00:00:23.920 Exactly, and sometimes, let's be honest, it leads to those 8 00:00:23.960 --> 00:00:27.519 glitches that are just hilariously memorable. 9 00:00:27.600 --> 00:00:30.079 Yeah, it's that invisible layer, isn't it, ye, shaping so 10 00:00:30.160 --> 00:00:33.399 much of how you experience the game, making it feel reactive, 11 00:00:33.560 --> 00:00:34.880 not just static. 12 00:00:34.600 --> 00:00:37.719 Totally, and to really get our heads around how it 13 00:00:37.759 --> 00:00:40.359 all works. We've gathered a whole stack of material. We're 14 00:00:40.359 --> 00:00:44.159 talking algorithms, techniques, some history. It's pretty deep stuff. 15 00:00:44.200 --> 00:00:45.039 You can get technical. 16 00:00:45.159 --> 00:00:47.799 Yeah, but our mission here for this deep dive is 17 00:00:47.840 --> 00:00:49.719 to sort of cut through that jargon. We want to 18 00:00:49.759 --> 00:00:53.359 find the really cool, important ideas and give you listening 19 00:00:53.399 --> 00:00:56.479 now a short tut to understanding what's going on behind. 20 00:00:56.200 --> 00:00:58.719 The scenes, like lifting the curtain a bit, Yeah. 21 00:00:58.799 --> 00:01:02.000 Uncover some surprises, maybe have a few aha moments along 22 00:01:02.000 --> 00:01:02.320 the way. 23 00:01:02.560 --> 00:01:04.879 So think of it like getting a peek under the hood, 24 00:01:05.680 --> 00:01:09.599 looking at the simulated brains, making everything tick, your opponents, 25 00:01:09.680 --> 00:01:12.519 your allies, even the whole world sometimes. 26 00:01:12.640 --> 00:01:15.120 Okay, so where do we start. The sources actually bring 27 00:01:15.200 --> 00:01:17.519 up an interesting point about the history. Back in the 28 00:01:17.560 --> 00:01:22.599 mid nineties, AI was actually like a selling point on 29 00:01:22.680 --> 00:01:23.519 the back of the box. 30 00:01:23.599 --> 00:01:26.040 Oh yeah, I remember that era. Beneath a Steel Sky 31 00:01:26.159 --> 00:01:27.280 is a good example. 32 00:01:27.000 --> 00:01:30.719 Right, It promised virtual theater. What was that supposed to be? 33 00:01:30.879 --> 00:01:34.480 Well, the idea, the ambition was for characters who weren't 34 00:01:34.519 --> 00:01:37.840 just puppets. They'd move around, do their own thing, make 35 00:01:37.879 --> 00:01:41.959 the world feel lived in. Virtual theater. Sounded grand, but 36 00:01:42.000 --> 00:01:44.480 the reality, the reality is the sources kind of point 37 00:01:44.480 --> 00:01:47.879 out was maybe a bit less revolutionary. Often it just 38 00:01:47.920 --> 00:01:51.519 met characters pacing back and forth on pretty fixed paths. Still, 39 00:01:51.560 --> 00:01:52.400 it was an early. 40 00:01:52.159 --> 00:01:55.599 Attempt, a bit underwhelming then maybe, But then something came 41 00:01:55.640 --> 00:01:58.640 along that really shifted things, didn't it, gold enneye double 42 00:01:58.640 --> 00:01:59.120 O seven. 43 00:02:00.480 --> 00:02:02.200 Yeah, that was definitely a game changer for a lot 44 00:02:02.239 --> 00:02:03.480 of people's perception of AI. 45 00:02:03.680 --> 00:02:04.640 What did it do differently? 46 00:02:04.959 --> 00:02:07.359 It wasn't just pre programmed patterns. I mean, yes, the 47 00:02:07.359 --> 00:02:10.199 guards us simple state machines, that's true, But the key 48 00:02:10.240 --> 00:02:12.240 thing was this basic sense simulation. 49 00:02:12.319 --> 00:02:13.919 They added sense simulation. 50 00:02:14.039 --> 00:02:17.000 Yeah, characters could essentially see their buddies. If one guard 51 00:02:17.039 --> 00:02:21.319 saw another get taken out, they'd react, maybe investigate, maybe 52 00:02:21.360 --> 00:02:24.800 raise the alarm. It wasn't super complex, but that idea 53 00:02:24.840 --> 00:02:28.400 of awareness of perceiving the world a change things big time, 54 00:02:28.599 --> 00:02:32.639 and that whole concept simulating senses, site, sound, it became 55 00:02:32.719 --> 00:02:35.080 absolutely central to stealth games. 56 00:02:34.840 --> 00:02:38.280 Later on, like Thief or Metal Here sawid. 57 00:02:38.120 --> 00:02:42.840 Exactly Thief, Metal Gear, splinter Cell. The core gameplay loop 58 00:02:43.039 --> 00:02:47.120 often revolves around exploiting or avoiding the AI senses how 59 00:02:47.159 --> 00:02:50.080 guards react to a noise or a shadow or seeing 60 00:02:50.159 --> 00:02:51.240 something suspicious. 61 00:02:51.479 --> 00:02:54.120 So this points to a fundamental cycle that seems to 62 00:02:54.199 --> 00:02:55.759 run through all the material we looked. 63 00:02:55.599 --> 00:02:57.439 At pretty much. Yeah, you can boil a lot of 64 00:02:57.439 --> 00:03:00.400 it down to perception leads to decision making, which leads 65 00:03:00.400 --> 00:03:00.800 to action. 66 00:03:00.919 --> 00:03:01.960 Per section decision action. 67 00:03:02.120 --> 00:03:05.280 Okay, the AI gathers info for the world, decides what 68 00:03:05.319 --> 00:03:09.120 to do about it, and then actually does something, move, shoots, hides, talks. 69 00:03:08.840 --> 00:03:12.280 Whatever, right, and to actually do those actions. Especially back then, 70 00:03:12.680 --> 00:03:15.520 developers were always fighting against hardware limits, weren't they. 71 00:03:15.639 --> 00:03:19.639 Oh, absolutely huge constraint. Early days, you had maybe one 72 00:03:19.639 --> 00:03:23.240 CPU doing everything. Now we've got multiple cores, powerful GPUs. 73 00:03:23.280 --> 00:03:25.080 PPUs are more for graphics though. 74 00:03:24.919 --> 00:03:27.840 Mostly Yeah, they're built for massively parallel stuff, tons of 75 00:03:27.840 --> 00:03:31.159 simple calculations at once, which is perfect for graphics, but 76 00:03:31.520 --> 00:03:34.240 it does influence AI. Sometimes. If you can break an 77 00:03:34.240 --> 00:03:38.080 AI task into lots of small independent bits, maybe you 78 00:03:38.120 --> 00:03:41.520 can leverage the GPU. But CPUs are still king for 79 00:03:41.560 --> 00:03:44.800 that complex, sequential thinking part of AI. 80 00:03:45.039 --> 00:03:48.240 So it's always a trade off performance versus complexity. 81 00:03:47.759 --> 00:03:49.639 Always has been, always will being gamed of. 82 00:03:49.800 --> 00:03:55.080 And it's not just consoles and PCs anymore. You've got phones, tablets, iOS, Android, 83 00:03:55.439 --> 00:03:56.800 AI has to run everywhere. 84 00:03:56.840 --> 00:03:59.680 Now, Yeah, the rise of mobile is huge, and that's 85 00:03:59.680 --> 00:04:03.199 where engines like Unreal, Unity, godot are so important. They 86 00:04:03.199 --> 00:04:05.319 provide this layer of this framework. 87 00:04:04.919 --> 00:04:06.439 So you can write the AI once. 88 00:04:06.479 --> 00:04:11.520 Pretty much, write at once, deploy it on PC, console, mobile, 89 00:04:12.680 --> 00:04:15.879 maybe with some tweaks, but the core logic can often 90 00:04:15.919 --> 00:04:18.720 be reused. That definitely shapes how you design the AI 91 00:04:18.759 --> 00:04:19.800 systems in the first place. 92 00:04:19.839 --> 00:04:22.879 Okay, so keeping that cycle in mind, perception, decision, action, 93 00:04:23.000 --> 00:04:27.000 and let's talk about the action bit, specifically movement. How 94 00:04:27.040 --> 00:04:31.360 do AI characters actually you know, get around because unless 95 00:04:31.360 --> 00:04:33.839 it's maybe an economic sim they need to move. 96 00:04:34.040 --> 00:04:38.079 Movement's fundamental. Yeah, an AI character has a position, maybe 97 00:04:38.079 --> 00:04:42.000 an orientation, physical properties. Yeah, and the movement algorithms figure 98 00:04:42.040 --> 00:04:44.800 out frame by frame where they need to go next 99 00:04:44.879 --> 00:04:45.600 to reach their goal. 100 00:04:45.639 --> 00:04:47.879 But you don't want them just snapping between locations. Right, 101 00:04:48.000 --> 00:04:50.040 has to look physical exactly. 102 00:04:50.120 --> 00:04:54.199 That's where kinematics comes in. It's basically applying physics Newton's laws. Really, 103 00:04:54.240 --> 00:04:58.959 even in a simple way. Velocity can't change instantly right inertia. Yeah, 104 00:04:59.040 --> 00:05:01.959 So the algorithms don't teleport the character. They consider its 105 00:05:02.000 --> 00:05:04.839 current speed and direction and use acceleration to make the 106 00:05:04.839 --> 00:05:08.120 movement smooth and believable. No instance stops or starts. 107 00:05:08.319 --> 00:05:11.240 And the sources talk about these things called steering behaviors 108 00:05:11.279 --> 00:05:12.680 like building blocks for a movement. 109 00:05:13.000 --> 00:05:17.120 They are think of them as simple reactive urges. Seek 110 00:05:17.279 --> 00:05:20.319 tells a character to move towards a target. Flea tells 111 00:05:20.319 --> 00:05:23.360 it to run away. Very basic, but you build on those, 112 00:05:23.439 --> 00:05:27.319 you combine them. So pursue isn't just seek. It tries 113 00:05:27.360 --> 00:05:29.480 to predict where a moving target will be and seeks 114 00:05:29.519 --> 00:05:33.959 that point. Collision avoidant steers away from obstacles. Often these 115 00:05:34.000 --> 00:05:37.720 more complex behaviors work by delegating to the simpler ones 116 00:05:38.360 --> 00:05:41.120 seek the target, but also flee from that wall that's 117 00:05:41.160 --> 00:05:41.920 getting too close. 118 00:05:42.160 --> 00:05:45.800 The material mentioned a potential issue with a simple wander behavior, 119 00:05:45.800 --> 00:05:48.519 something about rotational jerkiness. What does that look like? 120 00:05:48.800 --> 00:05:52.079 Imagine a character just randomly picking a new direction every 121 00:05:52.120 --> 00:05:54.959 single frame. They kind of twitch and spin around unnaturally. 122 00:05:54.959 --> 00:05:58.519 It looks terrible. Uh so simple wander needs smoothing. Maybe 123 00:05:58.560 --> 00:06:01.199 you pick a random point on a circle slightly ahead 124 00:06:01.240 --> 00:06:03.639 of the character and steer towards that. Updating it less 125 00:06:03.639 --> 00:06:06.199 frequently makes it look more like they're meandering, not having 126 00:06:06.199 --> 00:06:06.680 a seizure. 127 00:06:06.839 --> 00:06:10.959 Gotcha. So steering handles the how of moving smoothly, But 128 00:06:11.040 --> 00:06:14.680 how does the AI figure out the where? Especially in 129 00:06:14.720 --> 00:06:17.319 a big complex level. That sounds like pathfinding. 130 00:06:17.680 --> 00:06:21.000 That's exactly what pathfinding is for finding a route a 131 00:06:21.040 --> 00:06:23.720 sequence of steps from point A to point B that 132 00:06:23.800 --> 00:06:27.800 avoids opticles like walls, pits, whatever the level design are 133 00:06:27.800 --> 00:06:28.399 put in the way. 134 00:06:28.600 --> 00:06:31.720 It's the how do I get there from here? Problem? 135 00:06:31.800 --> 00:06:36.240 Precisely the source of mentioned classic algorithms Dykstra's algorithm. I 136 00:06:36.240 --> 00:06:38.560 think I remember that from a computer science class. 137 00:06:38.680 --> 00:06:42.160 Yeah, dykstras is fundamental graph theory, it's guaranteed to find 138 00:06:42.439 --> 00:06:46.000 the absolute shortest path from your start point to every 139 00:06:46.040 --> 00:06:48.720 other reachable point in the graph, representing. 140 00:06:48.199 --> 00:06:49.519 The level every other point. 141 00:06:49.560 --> 00:06:52.000 That sounds like a lot, it can be, and that's 142 00:06:52.040 --> 00:06:54.959 the catch. For games. Usually you only care about the 143 00:06:55.000 --> 00:06:58.639 past to one specific goal. Calculating the shortest path to 144 00:06:58.720 --> 00:07:00.959 everywhere else is often effort. 145 00:07:00.800 --> 00:07:03.480 So games tend to use a more A. 146 00:07:03.560 --> 00:07:05.879 Star A is generally the go to. Yeah. Yeah, it's 147 00:07:05.920 --> 00:07:09.439 smarter about searching, how so it uses a heuristic think 148 00:07:09.480 --> 00:07:11.480 of it as an educated guess or a rule of 149 00:07:11.519 --> 00:07:13.959 thumb to estimate the distance remaining to. 150 00:07:13.920 --> 00:07:16.519 The goal, like as the crow flies distance. 151 00:07:16.439 --> 00:07:21.360 Often yeah, or Manhattan distance. In grid based worlds, Igelegan 152 00:07:21.399 --> 00:07:24.959 combines the actual cost traveled so far with this estimated 153 00:07:25.000 --> 00:07:29.079 future cost. This lets to prioritize exploring paths that seem 154 00:07:29.120 --> 00:07:30.920 to be heading more directly towards the goal. 155 00:07:31.000 --> 00:07:33.600 Ah, so it doesn't waste time exploring paths that are 156 00:07:33.639 --> 00:07:36.600 clearly going the wrong way, even if they look short initially. 157 00:07:36.720 --> 00:07:40.120 Exactly, it focuses the search, makes it much faster than 158 00:07:40.240 --> 00:07:43.639 Diikstra in most typical game scenarios. Now there's a small 159 00:07:43.680 --> 00:07:49.279 trade off mentioned. Oh, if you're heuristic, your guess is bad. Specifically, 160 00:07:49.920 --> 00:07:53.519 if it overestimates the actual cost to the goal, then 161 00:07:53.600 --> 00:07:58.040 a might not find the absolute mathematically shortest path. It 162 00:07:58.120 --> 00:08:00.519 might find a path that looks shorter based on the 163 00:08:00.519 --> 00:08:01.120 bad gas. 164 00:08:01.199 --> 00:08:03.839 So developers might accept a path that's like ninety nine 165 00:08:03.839 --> 00:08:06.399 percent optimal if it means the calculation is way faster. 166 00:08:06.600 --> 00:08:10.399 That's the aha moment right there, exactly. Performance often trump's 167 00:08:10.399 --> 00:08:13.639 perfect optimality. In real time games, good enough, fast enough 168 00:08:13.839 --> 00:08:14.360 is often the. 169 00:08:14.319 --> 00:08:16.959 Mantra makes sense. So for A or diestra to even work, 170 00:08:17.000 --> 00:08:19.000 they need a map, right, but not the full three 171 00:08:19.079 --> 00:08:20.560 D graphics geometry. 172 00:08:20.160 --> 00:08:22.800 Where they need a simplified representation of the navigable space. 173 00:08:22.839 --> 00:08:24.879 You can't search through raw triangles. 174 00:08:24.560 --> 00:08:26.279 Like navigation meshes. I've heard that. 175 00:08:26.319 --> 00:08:29.800 Term nav meshes are super common. They basically represent all 176 00:08:29.800 --> 00:08:33.159 the walkable surfaces as a set of connected polygons like 177 00:08:33.240 --> 00:08:36.960 piles on the floor. The AI path finds across these polygons. 178 00:08:37.000 --> 00:08:39.639 How does that work with the algorithm? Are the polygons 179 00:08:39.679 --> 00:08:40.159 the nodes? 180 00:08:40.480 --> 00:08:42.759 Sometimes the center of the polygons, but often it's the 181 00:08:42.919 --> 00:08:46.720 edges between the polygons. The edges as nodes approach is mentioned. 182 00:08:46.879 --> 00:08:50.039 Path finding becomes finding a sequence of edges to cross 183 00:08:50.200 --> 00:08:52.639 to get from the starting polygon to the goal polygon. 184 00:08:52.799 --> 00:08:56.759 Okay, but what about huge open worlds. Those nav meshes 185 00:08:56.840 --> 00:08:57.879 must get enormous. 186 00:08:57.919 --> 00:09:01.440 They do, and that's where hierarchy pathfinding comes into play. 187 00:09:01.559 --> 00:09:04.480 Layers kind of you group areas together, like maybe this 188 00:09:04.519 --> 00:09:07.200 whole building is one high level note and that neighborhood 189 00:09:07.279 --> 00:09:10.440 is another. The AI first finds a path between these 190 00:09:10.480 --> 00:09:14.440 big regions, like planning a trip using major highways. Then 191 00:09:14.720 --> 00:09:17.320 once it enters a region, it does a more detailed 192 00:09:17.320 --> 00:09:20.759 pathfind within that smaller area, figuring out the local streets. 193 00:09:21.120 --> 00:09:24.639 That seems way more manutable. Okay, so characters can move 194 00:09:24.679 --> 00:09:27.720 realistically with steering, and they can find paths with pathfinding, 195 00:09:27.919 --> 00:09:31.159 But how do they decide decide where to go, whether 196 00:09:31.240 --> 00:09:34.440 to fight or hide or talk. That's the decision making 197 00:09:34.480 --> 00:09:34.879 part of the. 198 00:09:34.879 --> 00:09:37.440 Cycle, right, This is the core intelligence, the brain making 199 00:09:37.440 --> 00:09:40.159 the choices. It determines what the character intends to do. 200 00:09:40.320 --> 00:09:43.559 The sources mentioned decision trees as a common technique, simple 201 00:09:43.679 --> 00:09:44.399 if thin stuff. 202 00:09:44.799 --> 00:09:47.679 Pretty much, they're like flow charts, a series of questions 203 00:09:47.759 --> 00:09:50.679 usually yes no, leading down branches to an action is 204 00:09:50.720 --> 00:09:53.240 the enemy visible? Yes? Do I have AMMO? Yes? 205 00:09:53.279 --> 00:09:53.639 Shoot? 206 00:09:53.840 --> 00:09:56.799 No reload back up the tree is the endy visible? 207 00:09:57.039 --> 00:09:58.000 No patrol. 208 00:09:58.240 --> 00:10:00.639 They sound fast and easy for designers to work with. 209 00:10:00.759 --> 00:10:04.360 They are very fast, execute, modular, easy to build and tweak, 210 00:10:04.799 --> 00:10:07.480 use for tons of stuff from simple reactions to choose 211 00:10:07.480 --> 00:10:08.159 the animations. 212 00:10:08.240 --> 00:10:10.360 What about state machines those seem classic too. 213 00:10:10.679 --> 00:10:16.039 Finite state machines yeah, FSMs. You define distinct states. The 214 00:10:16.080 --> 00:10:21.120 AI can be in patrolling, investigating, attacking, fleeing, and transitions 215 00:10:21.120 --> 00:10:24.080 between them. If you're patrolling and you hear a noise, 216 00:10:24.480 --> 00:10:27.840 the transition condition you switch to the investigating state. 217 00:10:28.120 --> 00:10:30.480 Sounds neat, but maybe it gets messy. If you have 218 00:10:30.519 --> 00:10:31.720 lots of states. 219 00:10:31.440 --> 00:10:33.799 It can become a spaghetti diagram really quickly if you're 220 00:10:33.799 --> 00:10:37.000 not careful, lots of arrows criss crossing. That's why higer 221 00:10:37.039 --> 00:10:38.240 Arsenill state machines. 222 00:10:37.919 --> 00:10:40.360 Were developed, so states within states exactly. 223 00:10:40.720 --> 00:10:45.000 Your combat state might contain substates like taking cover, firing, reloading. 224 00:10:45.200 --> 00:10:48.519 It helps organize complexity, and you can have transitions that 225 00:10:48.600 --> 00:10:51.639 jump across levels, like a sudden loud noise pulling you 226 00:10:51.679 --> 00:10:54.600 out of reloading right back to a top level alert state. 227 00:10:54.840 --> 00:10:58.840 Now The material really highlights behavior trees as being super popular. Now, 228 00:10:59.080 --> 00:11:00.080 what's the deal with them? 229 00:11:00.559 --> 00:11:04.840 Behavior trees are very powerful and crusely very intuitive for designers. 230 00:11:05.399 --> 00:11:08.320 They structure tasks in a tree like way. The AI 231 00:11:08.440 --> 00:11:12.120 executes the tree usually depth first, trying branches until it 232 00:11:12.159 --> 00:11:14.399 finds an action or sequence that succeeds. 233 00:11:14.480 --> 00:11:15.440 Can you give an example? 234 00:11:15.559 --> 00:11:19.120 Sure, a root node might be engage enemy. It has 235 00:11:19.200 --> 00:11:22.639 children like is enemy visible? If yes, it goes down 236 00:11:22.679 --> 00:11:24.919 the attack branch. If no, it tries the search for 237 00:11:25.039 --> 00:11:27.759 enemy branch, which might have subtasks like go to last 238 00:11:27.799 --> 00:11:31.159 known position, then scan area. It feels logical, like how 239 00:11:31.159 --> 00:11:32.600 a person might approach a problem. 240 00:11:32.679 --> 00:11:35.080 Again, they're reusable, hugely reusable. 241 00:11:35.240 --> 00:11:37.960 You create nodes for common actions, move to play, animation, 242 00:11:38.159 --> 00:11:40.679 check condition, and assemble them into trees. You can make 243 00:11:40.679 --> 00:11:43.679 a complex take cover behavior and just drop it into 244 00:11:43.679 --> 00:11:47.879 different AI characters. Trees engines often have visual editors like 245 00:11:48.000 --> 00:11:48.519 Lego for. 246 00:11:48.519 --> 00:11:49.879 AI like al prefabs. 247 00:11:49.919 --> 00:11:51.759 Okay, that makes sense exactly that concept. 248 00:11:51.960 --> 00:11:54.720 Are there other ways? The sources mention things like utility 249 00:11:54.759 --> 00:11:57.639 based systems. Yeah, goal oriented behavior. 250 00:11:57.279 --> 00:12:00.000 Yeah, those are a different philosophy instead of pre defined 251 00:12:00.159 --> 00:12:06.360 logic flows. The AI has goals or motives hunger, thirst, safety, curiosity, 252 00:12:06.639 --> 00:12:09.080 each with a current insistence level. 253 00:12:08.879 --> 00:12:09.720 How much it wants it. 254 00:12:10.039 --> 00:12:12.440 Basically, yeah, the AI looks at all the actions it 255 00:12:12.440 --> 00:12:15.120 could do and evaluates how well each action would satisfy 256 00:12:15.200 --> 00:12:18.279 its most insistent needs. It picks the action that leads 257 00:12:18.279 --> 00:12:20.240 to the lowest overall discontentment. 258 00:12:20.480 --> 00:12:24.120 Sounds more adaptive, maybe less predictable, it can be. 259 00:12:24.559 --> 00:12:27.879 The big challenge is often calculating the utility of an action. 260 00:12:29.000 --> 00:12:32.720 How much will eating this apple reduce my hunger? How 261 00:12:32.759 --> 00:12:35.639 long will it take to pathfind to the apple? Often 262 00:12:35.679 --> 00:12:38.399 requires good estimation heuristics. 263 00:12:37.799 --> 00:12:39.799 Again, seems complex to set up right. 264 00:12:40.000 --> 00:12:42.080 It can be more work up front than say, a 265 00:12:42.159 --> 00:12:46.159 state machine. Rule based systems are another alternative mentioned, using 266 00:12:46.399 --> 00:12:49.600 if then rules like if player is visible and health 267 00:12:49.679 --> 00:12:53.840 is low, then find cover. But writing good comprehensive rule 268 00:12:53.879 --> 00:12:55.039 sets is really hard. 269 00:12:55.159 --> 00:12:57.960 The knowledge acquisition bottleneck they called it. 270 00:12:57.919 --> 00:13:00.879 Right, So while powerful in theory, they're less common in 271 00:13:00.919 --> 00:13:04.480 practice for moment to moment character AI compared to trees 272 00:13:04.600 --> 00:13:05.320 or state machines. 273 00:13:05.399 --> 00:13:09.559 Okay, individual decisions covered. What about coordinating groups or making 274 00:13:09.600 --> 00:13:10.879 bigger strategic plans? 275 00:13:10.919 --> 00:13:14.039 Yeah, moving beyond single characters, that's tactical and strategic. AI 276 00:13:14.559 --> 00:13:18.519 tactical is about local coordination soldiers using cover fire, flanking maneuvers. 277 00:13:19.000 --> 00:13:22.120 Strategic is the big picture, deciding which continent to invade 278 00:13:22.159 --> 00:13:25.159 in a global strategy game, or managing an economy in 279 00:13:25.200 --> 00:13:26.120 an RTS. 280 00:13:26.159 --> 00:13:30.320 For tactics, the sources mention tactical waypoints, not just points 281 00:13:30.360 --> 00:13:33.240 to walk to, but points with extra info. 282 00:13:33.480 --> 00:13:36.759 Exactly, a waypoint might be tagged as good cover spot 283 00:13:36.879 --> 00:13:40.879 versus north or ambush point, but just relying on static 284 00:13:40.879 --> 00:13:44.000 points placed by a designer has limits. What if the 285 00:13:44.039 --> 00:13:45.519 situation changes. 286 00:13:45.320 --> 00:13:47.320 So run time analysis. 287 00:13:46.879 --> 00:13:51.080 Yeah, better systems analyze the environment dynamically, calculating lines of sight, 288 00:13:51.399 --> 00:13:54.120 checking actual cover effectiveness based on where the enemies are 289 00:13:54.200 --> 00:13:54.639 right now. 290 00:13:54.879 --> 00:13:58.639 The frag maps idea sounded cool learning from where players die. 291 00:13:58.879 --> 00:14:02.480 It's a neat example of simple tactical learning. Basically, areas 292 00:14:02.480 --> 00:14:04.919 on the map build up a danger value. If AI 293 00:14:05.000 --> 00:14:08.120 units keep getting killed there, future AI might then learn 294 00:14:08.159 --> 00:14:10.159 to avoid those spots or be more cautious. 295 00:14:10.200 --> 00:14:12.120 But they need to unlearn it too, right, so they 296 00:14:12.159 --> 00:14:15.039 don't avoid a place forever after one unlucky death. 297 00:14:15.200 --> 00:14:18.519 Exactly, a decay factor or unlearning bias is important and. 298 00:14:18.440 --> 00:14:21.279 For strategy, Like in an RTS, how does the AI 299 00:14:21.320 --> 00:14:23.919 decide whether to build more units or research tech. 300 00:14:24.120 --> 00:14:27.639 Often with multi tiered AI, you have layers. A top 301 00:14:27.759 --> 00:14:30.960 level strategic layer makes broad decisions we need more resource 302 00:14:31.000 --> 00:14:33.279 income or prepare for an attack on their. 303 00:14:33.200 --> 00:14:35.200 Base, and that gets passed down right. 304 00:14:35.279 --> 00:14:38.399 A mid level tactical layer might translate attack their base 305 00:14:38.639 --> 00:14:42.519 into a simple strikeforce alpha A coordinates x y and 306 00:14:42.559 --> 00:14:46.600 proceed along route Z. Then the lowest level AI, the 307 00:14:46.639 --> 00:14:50.519 individual unit AI, handles the actual movement in shooting based 308 00:14:50.559 --> 00:14:51.279 on those orders. 309 00:14:51.559 --> 00:14:54.360 So the player is often the top strategic layer in 310 00:14:54.399 --> 00:14:55.080 an RTS. 311 00:14:55.200 --> 00:14:58.080 In many cases, yes, the player sets the goals and 312 00:14:58.200 --> 00:15:01.080 the AI executes the details, though some games have high 313 00:15:01.120 --> 00:15:03.679 level AI opponents making strategic choices. 314 00:15:03.360 --> 00:15:06.879 Too, and tactical pathfinding came up again. Path cost isn't 315 00:15:06.919 --> 00:15:08.200 just distance, right. 316 00:15:08.279 --> 00:15:11.080 It's about the quality of the path caxically speaking. Is 317 00:15:11.120 --> 00:15:14.639 it exposed to enemy fire? Does it offer good ambush opportunities? 318 00:15:15.000 --> 00:15:17.879 A path might be longer but much safer, so tactical 319 00:15:17.919 --> 00:15:19.039 pathfinding would prefer it. 320 00:15:19.159 --> 00:15:23.200 Okay, The big one learning Can game AI actually learn 321 00:15:23.360 --> 00:15:24.799 like you know? Learn learn? 322 00:15:25.320 --> 00:15:28.519 The sources definitely cover learning, but with a crucial caveat 323 00:15:28.559 --> 00:15:31.960 for games, the goal isn't just effective AI. It's AI 324 00:15:32.039 --> 00:15:34.799 that results in good gameplay. An AI that learns an 325 00:15:34.879 --> 00:15:36.639 unbeatable exploit isn't fun. 326 00:15:36.919 --> 00:15:39.240 That's a really important distinction. So what kinds of learning 327 00:15:39.240 --> 00:15:39.799 do they use? 328 00:15:40.039 --> 00:15:44.240 Some simple stuff like parameter modification basically tweaking numbers like 329 00:15:44.519 --> 00:15:47.960 if the AI keeps losing, maybe slightly increase its accuracy. 330 00:15:48.000 --> 00:15:52.879 Parameter uses optimization techniques like kill climbing. Then there's action prediction, 331 00:15:53.120 --> 00:15:55.360 trying to guess what the player will do next based 332 00:15:55.399 --> 00:15:58.279 on past behavior using techniques like n grams. 333 00:15:58.000 --> 00:16:01.639 And reinforcement learning. That's the hot top AI. Generally trial 334 00:16:01.639 --> 00:16:02.480 and error. 335 00:16:02.279 --> 00:16:06.440 Exactly the AI tries ashs, gets rewards or penalties from 336 00:16:06.480 --> 00:16:09.919 the game, find treasure, reward, fall in lava, penalty, and 337 00:16:10.039 --> 00:16:13.039 learns over time which actions lead to good outcomes in 338 00:16:13.080 --> 00:16:16.960 different situations or states. Q learning is a common example mentioned. 339 00:16:17.000 --> 00:16:18.360 What's the challenge there? For games? 340 00:16:18.799 --> 00:16:23.360 Defining the state? Real game worlds are incredibly complex and continuous. 341 00:16:23.879 --> 00:16:27.879 Trying to break down every possible situation, player, position, enemy health, 342 00:16:28.039 --> 00:16:31.120 ammial count, time of day into discrete states for the 343 00:16:31.159 --> 00:16:32.559 algorithm can be really. 344 00:16:32.320 --> 00:16:34.799 Hard and neural networks. Deep learning. 345 00:16:35.120 --> 00:16:38.799 Neural nets are powerful pattern recognizers. They learn by adjusting 346 00:16:38.840 --> 00:16:42.639 connections between artificial neurons. They were used historically in things 347 00:16:42.679 --> 00:16:45.960 like checkers programs or td GAMIN for evaluating board positions. 348 00:16:46.279 --> 00:16:49.879 Deep blue used complex evaluation, possibly including n ends. 349 00:16:50.000 --> 00:16:52.519 But are they used for deciding if a guard should 350 00:16:52.519 --> 00:16:53.120 open a door? 351 00:16:53.720 --> 00:16:57.879 Less so, according to the sources, for general character behavior logic, 352 00:16:58.200 --> 00:17:00.919 things like behavior trees are often more press tactical, easier 353 00:17:00.960 --> 00:17:04.279 to design, debug and control. Neural nets are great for 354 00:17:04.359 --> 00:17:08.440 specific tasks like image recognition or maybe complex evaluation, but 355 00:17:08.599 --> 00:17:11.440 maybe overkill or too inch of a black box for 356 00:17:11.559 --> 00:17:13.599 typical moment to moment decisions. 357 00:17:13.200 --> 00:17:16.160 Right now, So, traditional methods are still really dominant for 358 00:17:16.240 --> 00:17:18.119 a lot of character AI for. 359 00:17:18.079 --> 00:17:21.599 The actual decision logic very much so. The complexity, data 360 00:17:21.640 --> 00:17:25.200 needs and training time for deep learning for general flexible 361 00:17:25.279 --> 00:17:28.839 character behavior are still pretty significant hurdles compared to more 362 00:17:28.960 --> 00:17:30.000 established techniques. 363 00:17:30.279 --> 00:17:35.759 Okay, wow, so we've got movement, pathfinding, decisions, tactics, strategy, 364 00:17:35.799 --> 00:17:38.920 even learning. How does all this stuff actually run inside 365 00:17:39.000 --> 00:17:42.160 the game, and how does the AI even know what's 366 00:17:42.200 --> 00:17:43.559 happening in the world. This is the. 367 00:17:43.480 --> 00:17:48.839 Plumbing right exactly, The implementation details Yeah, super important. First off, scheduling, 368 00:17:49.720 --> 00:17:53.400 you might have hundreds of AI characters. You can't afford 369 00:17:53.440 --> 00:17:57.119 to run every single AI calculation for every character every 370 00:17:57.119 --> 00:17:59.960 single frame. Your game would grind to a halt. 371 00:18:00.240 --> 00:18:02.240 So you need to manage who thinks. 372 00:18:02.000 --> 00:18:06.640 When precisely you schedule AI updates. Maybe important enemies nearby 373 00:18:06.720 --> 00:18:09.680 update every frame, but background characters only update every few 374 00:18:09.720 --> 00:18:13.440 seconds or when something relevant happens. High archo of scheduling 375 00:18:13.599 --> 00:18:15.839 helps manage this complexity. 376 00:18:15.240 --> 00:18:16.960 And anytime algorithms what are they? 377 00:18:17.160 --> 00:18:19.720 Those clever algorithms design so you can interrupt them part 378 00:18:19.720 --> 00:18:21.920 way through their calculation and they can still give you 379 00:18:21.960 --> 00:18:23.880 a usable, even if not perfect result. 380 00:18:24.200 --> 00:18:27.039 Ah. So the AI doesn't just freeze for three seconds 381 00:18:27.039 --> 00:18:29.440 while it calculates the perfect path exactly. 382 00:18:29.599 --> 00:18:31.880 You can get a rough path, quickly, start moving, and 383 00:18:31.920 --> 00:18:33.839 maybe refine the path in the background if it gets 384 00:18:33.839 --> 00:18:37.559 more processing time later. Avoids that deer in headlights. 385 00:18:37.200 --> 00:18:40.440 Lack that can level of detail for AI similar to graphics. 386 00:18:40.160 --> 00:18:44.240 Yep, same principle. AI complexity scales with importance usually distance 387 00:18:44.839 --> 00:18:49.160 faraway AI runs simpler logic, closer AI runs the full simulation. 388 00:18:50.000 --> 00:18:53.440 Behavior Compression is mentioned as a way to smoothly transition 389 00:18:53.559 --> 00:18:55.599 between these levels. So the change is in jarring. 390 00:18:55.680 --> 00:18:58.079 Okay, how does the AI get its info? How does 391 00:18:58.079 --> 00:18:59.480 it perceive the world state? 392 00:19:00.079 --> 00:19:04.720 Good question. The simplest way is polling. The AI just 393 00:19:04.799 --> 00:19:06.880 asked the game can I see the player? Is there 394 00:19:06.880 --> 00:19:09.480 a wall in front of me? Whenever it needs to know? 395 00:19:09.720 --> 00:19:12.640 Sounds simple, but maybe slow if everyone's asking all the. 396 00:19:12.599 --> 00:19:15.960 Time, it doesn't scale well exactly. Imagine hundreds of AIS 397 00:19:16.079 --> 00:19:19.079 constantly querying the world state. Very inefficient. 398 00:19:19.200 --> 00:19:19.920 So what's better? 399 00:19:20.160 --> 00:19:22.759 The event passing is much more common in complex games. 400 00:19:23.079 --> 00:19:26.680 The game generates events sound created at location X, player 401 00:19:26.720 --> 00:19:29.559