need better AI for live ops, not just packaged product
no longer limited to moment-to-moment, short-term player actions data is ubiquitous: in-game, cross-game, related to real life (e.g. AR games, exergaming, Miiverse) -> enables AI to model long-term patterns in player behavior and develop a transferrable model of gameplay skills
industry: Steam, EA Origin, Zynga, Facebook, Kongregate, etc. modding: LB/LB2, SC/SC2, WoW addons, (modding book) able to continually A/B test and iterate on product in response to users -> enables AI to automatically modify game in response to ongoing development -> shifts from short-term NPC reactions or creating a full game to understanding operation of game and relationship to other games -> bigger emphasis on externalities of other games, transaction/payment methods
design and development is for second-screen; multi-platform; multi-game better tools means people can turn out single games well, but cross-game aspects are wild west -> AI systems should shift from game-specific to capable of handling different game settings; e.g. BTs are generic control for most avatar-bound NPCs, MDPs+utility models are recently becoming popular for generality -> from single player to communities of players (w/in and outside game) -> from characters w/in game to players across characters (and consumers)
reactive NPCs surrogate for opponents / partners
single/groups of NPCs interacting w/human on moment-to-moment basis alpha-beta search -> MCTS planning, BTs, HTNs fuzzy reasoning influence maps
game AI book?
PCG + authoring tools - ease creation player modeling w/in game - understand / visualize results (game analytics) AI as game design tool; AI agent as designer
ML to support analysis of games AI to generate content as tool or surrogate AI to adapt content drama management evolutionary computing (now MCTS?) asp + combinatorial search methods multi-agent systems (pons) machine learning (yu) [Q: game sketching tools? too heavy on existing?] – agustin;
yannakakis + togelius PCG overview, EDPCG yannakakis next wave paper bakkes on adaptive AI hendrickx on pcg overview togelius+schmidhuber; browne; cook(mech miner); mahlmann a smith ASP paper, VF, Ludocore; horswill+foged Nelson papers; Treanor Mateas on EAI, Facade, Prom Week g smith for mixed initiative DM overviews (Mateas 1999, Roberts+Isbell, Riedl+Bulitko) game analytics: medler, drachen, el nasr, game analytics book, wallner game analysis viz overview design experiments/optimization: lomas, RRT aiide, rafferty+kujala for optimal experimental design, jaffe, anderson (here or wave 3?) soc modeling[here or elsewhere?]: williams, …
cross-game representations for game systems (formal models) cross-game player models (not just per-game characters) –> differences b/t character, player, customer multi-game/cross-game generation methods AI agents that interface w/player across individual aspects (tutorials, sales agent?, long-term companion/pet) AI agents interfacing with player community proactive sensing, ARG, and interface b/t “RL” and game activities
generic game representations
formalisms: Jarvinen, Cook (skill atoms), Dormans, ebner, zagal+mateas bickmore for lifelong agents amershi + interactive ML? zeng? shivaswamy? settles? bengio for rep learning? nan li? king on robot scientist?
pcg generator generator? pereira for socailly present board game agents (w/in vs cross-game) crowdsource: shaker, boyang li, orkin player models: trueskill, huang halo, backrach+minka,
Q: WHERE DO GENERAL GAME PLAYING AGENTS FIT? – bellemare for environment genesereth description
reactive NPCs that data-mine appropriate responses long-term w/in game companions long-term meta-game companions
generating content across games -> e.g. Kongregate achievements in meta-game cross-game player modeling community modeling / matchmaking
– reponsive to player + game design models – acting to improve toward experience – acting to gain info on player for re-use
– continues to adapt + recall long-term history w/player – potentially a recurring character/partner across many games or contexts
– your guide to Miiverse – concept of drama manager taken to cross-game setting – [Q: why is this point game AI and not just AI?]
– automatically matching players to appropriate content – examples: ITSs w/exploratory domains proactive sensing – using players to learn about consumption habits, world
producer: game designer (tools) – Nelson, Treanor, Smith, EAI consumer: player (models) – game analytics (Drachen, Yannakakis, Medler), soc sci of games (Williams) developer: game design space learning/modeling
big data is useless/meaningless unless we find better ways to teach AI – users provide “demonstrations” ala Spore, NPCEdit, Mehta work – users provide subjective feedback ala Yu, Yannakakis – users tweak existing content ala Nickerson, (crowdsoruce) – interactive ML, GWAPs starting to explore, but how contextualize w/in game
logical conclusion of A/B testing and analytics is automated iteration early work should focus on minor changes and ways to intelligently search micro-games designed for particular user + goal examples: Cook, Browne, Schmidhuber, Hastings also use tools for authoring w/in system: Smith, Treanor, Nelson –> need to move beyond trying a single game to spanning games as a design space (game ecosystem)
(1) large virtual worlds (2) large sets of micro-games how do users discover what they want in the game? find new things they didn’t expect? examples: Yu,
users are content – social aspects of game (groups) are key users make content – active creation = Spore, Second Life, etc. – passive creation = Demon’s Souls examples: TrueSkill, ELO, (game analytics team ratings)
need to understand players, not just characters how transfer behavioral models b/t games? how model relationship b/t game features + users to adapt appropriately?
mentoring/tutoring AI to adapt structure of game to suit player lifecycle
matchmaking + recommendation to focus on high-value content w/in game
promotion of player socialization
cross-game player models