3 ms·
Don't know how many gamedevs are here but I've always wondered if within gamedev, there's differences in how much info/best practices are propagated within the
by danso 10y ago
Don't know how many gamedevs are here but I've always wondered if within gamedev, there's differences in how much info/best practices are propagated within the field. Graphics seems to be something that devs keep tabs on with each other. But maybe AI is more siloed off? I'm not sure what 10 years in improvement in AI looks like (as I do with graphics), but the algorithms and details [0] seem straightforward, or at least not impossible for a talented team with resources to build upon. I haven't played F.E.A.R. so maybe there's something about the game that vastly simplifies the factors?
[0] https://news.ycombinator.com/item?id=14029136 https://news.ycombinator.com/item?id=14029136
- buzzybee 10y agoAI is like rendering in that the programmers doing that tend to cluster together at conferences, so they aren't completely siloed. The subfields that don't have that clustering effect are the ones not established enough that it's become a known specialty. Where isolation tends to be more obvious is when looking between different sectors of gaming: people working on mobile F2P have a vastly different view from AAA console devs, and both are likewise different from indie PC devs. But AI in games does have a bit of an identity issue at times in that the obvious application of game AI(an AI opponent) doesn't represent the broader picture(FSMs appearing in every character controller and interactive element, planners employed in some forms of procedural content generation). With machine learning being the hot thing now, some AI devs are looking for angles to use it, which tends to push them towards backend data work that isn't too far removed from what other web and mobile companies are doing.
- douche 10y agoIn my admittedly limited exposure, game AI and academic/enterprise AI appear to be the most distant of cousins, at best.
- dkersten 10y agoThis is my understanding as well. Game AI is about "looking" intelligent (or as someone else in this thread said, about providing a fun experience to the player) and machine learning is about finding optimal solutions or predictions. Also, nobody wants to run a google-sized deep learning network or what have you in order to train a mere game AI.
- kevingadd 10y agoSecrecy is pervasive in the game industry, and used to be much worse. In many cases it goes as far as lawsuits over things that would be relatively minor and not worth suing over in other industries - like the look and feel of speech bubbles in multiplayer games. Things have gotten better over the last decade or so, with game studios letting developers publish research papers or slide decks or even entire software stacks as open source. Sometimes studios pay to send developers out to give talks and do recruiting. A decade ago, it would have been a hard sell to even get permission to buy your own ticket to a developer conference at most studios. The last AAA studio I worked for had similar policies - only a few people at the studio ever went to conferences, and we all had to pay out-of-pocket. That studio's policies have changed and they now let employees give talks. In comparison, the silicon valley firms I worked for had better wages and encouraged us to both attend and give talks. I think some of the improvements in games industry culture here have probably been triggered as a result of competition from SV startups and open source. When it comes to AI there are specific challenges: It's very hard to repurpose one game's AI for solving another game's challenges. It's very difficult to build truly general AI frameworks. Good game AI depends on knowledge of game systems, content, and mechanics - how each ability works, how the environments are laid out, the structure of level objectives, etc. All of these things can end up changing during development, and in a multiplayer game they may vary from match to match. As a result, even if you have a bunch of knowledge from other titles and you can reuse source code from the best titles in the market, you may still fail to build good AI for your game on your first few tries. It's that difficult. General purpose machine learning frameworks have helped here though, and people using off-the-shelf engines like Unreal can take advantage of all the existing tools there to start building AI quicker. So that's nice.
- johncolanduoni 10y agoIs the need to make your AI act as if they're working off of a different information and ability set than they are a factor? For example, I remember when you turned the difficulty up to 11 on the AI from the first Black Ops game, the enemies just turned into aimbots with a half second or so delay before firing (made watching the kill cam pretty funny). On the other hand, it's not feasible to hand your AI a depth field (which is already not super cheap to produce) and have them use real computer vision methods on it. Also it would probably be too strong a nerf, and now you'd have a very expensive and laughably inept AI.
- quadcore 10y agoGamedev here. First, remember the first principle of game AI, the goal is not to defeat the player but to entertain him. A simple way to say it is that too good AI is not what you want. Secondly, for most games, the fun is simply not on the AI component of the game. If you take those two points together, 1) good game AI is tricky concept, 2) that's not where the fun is anyway, then you got that not a lot of people cares about game AI.
- Insanity 10y agoFormer gaming-addict here. Does the fact that people tend to spend more time on the multiplayer component of a game rather than the single player play any role in this 'not caring about AI'. I mean, I really did not enjoy playing against AI because they are quite stupid in most cases, and the only way to make a singleplayer become harder does not seem to be to make the AI smarter, but rather have some 'modifiers' against the player. For example in L4D, the damage you take from a hit by a common zombie on normal difficulty is 10HP (out of 100HP), but on expert the same hit is worth 20HP. Other things to make it harder would be to spawn in more NPCs and just get the player more busy, or speed up the events in the game (like tetris). But in all games with a decent multiplayer, it is just _way_ more fun to play against other human beings rather than shooting a bot. Because it actually poses a challenge. EDIT: typo fix
- degenerate 10y agoI second this... the only reason I jump straight into the multiplayer when I try a new game is because the single player AI, at every difficulty level, in almost every game, is dumb as rocks. Making the AI "way too fucking smart" at insane/hard difficulty and then dumbing it down seems way, way, way better than hitpoint modifiers. I don't think smart AIs are written and then scrapped... that cannot possibly be what really happens?! Horrible if true.
- shrimp_emoji 10y ago>the only reason I jump straight into the multiplayer when I try a new game is because the single player AI, at every difficulty level, in almost every game, is dumb as rocks. There's more to it than that for me. No matter how good the AI is, shooting at AI feels empty and dull compared to shooting at actual people. Not only are you assured cunning, but also someone who hates your guts and lets you know in chat! I suspect many others feel that way too, deprecating singleplayer in favor of multiplayer from the get-go. >Making the AI "way too fucking smart" at insane/hard difficulty and then dumbing it down seems way, way, way better than hitpoint modifiers. It sure does. Have you ever played Civilization? The AI is always dumb; the only way to make the game harder is to make the game fundamentally unfairer. In this case, making the AI smarter is a much harder problem than in a shooter game though.
- 60654 10y agoSure, I can answer your question. Probably the largest community of game AI devs is the Game AI Programmers Guild [1], which organizes an AI summit at the GDC every year, and has its own mailing list etc. On the European side, the Nucl.ai conference [2] fulfills a similar need for devs who don't want to travel nine time zones to California. :). Finally on the academic side, the annual AIIDE conference [3] brings both academics and devs together. Between those three and their proceedings or conference videos, it's really easy to stay on top of developments in the industry. For an example, check out the AI Summit videos on the GDC Vault - many of them are free. Planning approaches were a pretty new thing in games 10-15 years ago, but since then there's been a lot of work on exploring them, and people understand their benefits and limitations much better these days. For an example, there's a great study of the use of planning across many games by Eric Jacopin at GDC 2015 AI summit. 1. http://www.gameai.com/ http://www.gameai.com/ 2. http://events.nucl.ai/ http://events.nucl.ai/ 3. https://sites.google.com/view/aiide2017/ https://sites.google.com/view/aiide2017/