Naughty Alien Posted July 2, 2012 Share Posted July 2, 2012 ..recently, I started some 'experiments' with some of ideas i have regarding procedural AI..what i mean is basically, capability that NPC's could 'learn' about world surrounding them, as well as property of such world..NPC's should also be able to 'forget' under specific circumstances, and so on..I did some base skeleton of such structure and im rather happy with it..however, i need to make a further move and start 'determining' way of learning..so, I was wondering, do you guys know about any good read, related to such topic, where i could gain some ideas how to properly do such thing, like making NPC 'learning'.. Quote Link to comment Share on other sites More sharing options...
Pixel Perfect Posted July 2, 2012 Share Posted July 2, 2012 There have certainly been many attempts at using neural nets in the past where NPCs, following feedback from humans, learn to navigate and interact with their game world. I believe they have had mixed results like most AI systems. Quote Intel Core i5 2.66 GHz, Asus P7P55D, 8Gb DDR3 RAM, GTX460 1Gb DDR5, Windows 7 (x64), LE Editor, GMax, 3DWS, UU3D Pro, Texture Maker Pro, Shader Map Pro. Development language: C/C++ Link to comment Share on other sites More sharing options...
Naughty Alien Posted July 2, 2012 Author Share Posted July 2, 2012 ..mm..well..Im trying just to figure out, best possible structure, in data terms, for 'learning curve', and not particularly simulation of humans..I have some limited success, but somethings missing (when my NPC 'unlearn' something, cant get it back if certain situation at initial state was actually making NPC to learn)..i must admit, its really hard, but not impossible..I just need some more read about this, but I cant find any single article talking about this or similar things.. Quote Link to comment Share on other sites More sharing options...
Pixel Perfect Posted July 2, 2012 Share Posted July 2, 2012 Well neural nets are a natural mechanism for storage of knowledge it's just, to be effective, they need to have something teach them (that is, indicate when they are right or wrong or at least some sort of encouraging or discouraging feedback system). Humans normally are best placed to do that but it doesn't preclude software from doing so if you have a good enough definition of the 'goal'. The problem with neural nets is you end up with something that works or semi works but no one really understands why they do so the downside is the inability to correct issues they may manifest. Making them unlearn something for example might prove impossible! The data structure in Neural Nets is effectively connections between nodes (neurons) which are continuously made broken and remade as the net learns. The knowledge is reflected in the eventual structure of the inter-wiring of the nodes. How one would simulate this by more conventional data structures I'm not at all sure. I don't have any experience in this area but maybe others do and will offer you some insight. Either way, good luck with this, it sounds fascinating Quote Intel Core i5 2.66 GHz, Asus P7P55D, 8Gb DDR3 RAM, GTX460 1Gb DDR5, Windows 7 (x64), LE Editor, GMax, 3DWS, UU3D Pro, Texture Maker Pro, Shader Map Pro. Development language: C/C++ Link to comment Share on other sites More sharing options...
Pixel Perfect Posted July 2, 2012 Share Posted July 2, 2012 Might be worth googling 'emergent AI' and see what that brings back Quote Intel Core i5 2.66 GHz, Asus P7P55D, 8Gb DDR3 RAM, GTX460 1Gb DDR5, Windows 7 (x64), LE Editor, GMax, 3DWS, UU3D Pro, Texture Maker Pro, Shader Map Pro. Development language: C/C++ Link to comment Share on other sites More sharing options...
Rick Posted July 2, 2012 Share Posted July 2, 2012 Could you do a pathfinding of sorts with nodes that have knowledge in them? Weighting certain nodes so the "right" answer is found. Then over time if not used node weights get reduced so it's possible the "wrong" answer is done. Quote Link to comment Share on other sites More sharing options...
flachdrache Posted July 3, 2012 Share Posted July 3, 2012 had been outsmarted by fuzzy logic because fuzzy defines what a agent/bot wants to goal in relationship of what is resonable in a certain situation - last thing i road however is that neural nets are considered practical in city-raceGames. quake2 is still the best solution to learn/test/debug AI ... link : http://botepidemic.no-origin.net/neuralbot/index.htm give`em pain & hf 2 Quote AMD 64 X2 Dual 5k - 4GB - XFX GForce9800GT - nv196.21 - WinXP Sp3 zBrush4R2 - Silo2Pro - Unwrap3DPro - Gile - MaPZone2.5 Xxploration FPS in progress ... Link to comment Share on other sites More sharing options...
Game Producer Posted July 3, 2012 Share Posted July 3, 2012 some AI books... that contain SOME info on procedural/learning site with some articles: http://www.ai-junkie.com/ book "Programming game AI by example" -> some perhaps relevant articles & fine book me thinks book "Artificial intelligence for games" http://www.amazon.com/Artificial-Intelligence-Games-Second-Edition/dp/0123747317/ another neat book, bit pricey but worth checking try keyword search "machine learning" and you might find some relevant stuff Quote Intel Dual Core 3GHz / GeForce GTS 450 - 1024 MB / Driver ver 267.59 / 8 GB RAM / Win 7 - 64 bit / LE2.50 / BMAX 1.48 game producer blog - Dead Wake Zombie Game powered by Leadwerks Engine Twitter - Unfollow me Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.