Game AI Pro 2: Collected Wisdom of Game AI Professionals

Source metadata

Type: Edited practitioner anthology
Editor: Steve Rabin
Year: 2015
Publisher: CRC Press (Taylor & Francis)
Section editors: Neil Kirby (General Wisdom), Alex Champandard (Architecture), Nathan R. Sturtevant (Movement & Pathfinding; Search), Damián Isla (Tactics & Spatial Awareness), Kevin Dill (Character Behaviour; Analytics, Content, Experience)

Forty-two chapters across seven sections. Sections I–V (General Wisdom, Architecture, Movement, Search, Tactics) are substantially covered by earlier Game AI Pro 360 anthology ingest. The novel material for this wiki is concentrated in Section VI (Character Behaviour) and Section VII (Analytics, Content Generation, and Experience Management).
Three chapters dissect AI from The Last of Us (Naughty Dog, 2013) at Infected/human-enemy/buddy levels — concrete implementation details for a landmark shipped game.
Sections VII introduces PCG, simulation-based world generation, analytics-driven adaptation, and AI-driven narrative experience management as a unified cluster of topics not previously covered in the wiki.
The Possibility Maps chapter (Section II) introduces a novel spatial-reasoning pattern for opportunistic AI not present in existing pages.
Smart Zones (Section II) describes a scene-management architecture for ambient NPC life explicitly implemented in Unity 3D.
Psychologically Plausible Methods for Character Behaviour Design (Section VI) grounds NPC animation and motion design in perception psychology (Johansson, Heider & Simmel, attribution theory).
Using Queues to Model a Merchant’s Inventory (Section VI) presents the M/M/1 queuing model as a computationally cheap way to simulate open-world economy without simulation.

“Character behaviour design is a visual communication problem: players parse motion through evolved shape-fitting mechanisms, and behaviour must be engineered so those mechanisms produce the intended reading.” (Carlisle, Ch. 38)
“The key idea [of AI experience management] is to create an AI game master (AI GM), which is bundled with every copy of the game … [that] monitors the player’s actions and dynamically modifies the story.” (Bulitko et al., Ch. 42)
“PCG can provide large amounts of content, so that each time the player starts the game, they will have a different experience … in combination with an AI system that can infer player skill, PCG can be used as a form of dynamic difficulty adjustment.” (Smith, Ch. 40)
“To make the world believable, it is sufficient to generate random inventories each time a merchant is encountered provided that they are consistent with players’ common sense expectations as to how they should change with time.” (Manslow, Ch. 37)
Possibility maps allow an AI to defer instantiation decisions until the player forces them, enabling the AI to choose the most opportune moment rather than committing to a position early. (Manslow, Ch. 7)
Dwarf Fortress’s four simulation principles: don’t overplan, break down the system, don’t overcomplicate, base the model on real-world analogs. (Adams, Ch. 41)

procedural-content-generation-ai (new page, this ingest)
ai-experience-management (new page, this ingest)
npc-behaviour-readability (new page, this ingest)
possibility-maps (new page, this ingest)
smart-zones (new page, this ingest)
economy-simulation-queues (new page, this ingest)
player-modelling (existing page — analytics-based modelling section is additive)
buddy-ai (existing page — Ellie chapter; already covered in earlier ingest)
history-of-game-ai (PCG and experience management timelines)
procedural-generation (existing game-design page — now has AI-specific companion)
game-analytics (existing page — analytics-based adaptation section is additive)
dwarf-fortress (existing notable — simulation principles now sourced here)

How do Smart Zones integrate with modern ECS architectures (DOTS/Unity)? The chapter prototype used a MonoBehaviour-style implementation; DOTS would require a different decomposition.
AI experience management systems like PaSSAGE and PACE were tested in relatively linear games. How do they scale to open-world games with very large narrative state spaces?
The merchant queue model assumes Markovian arrival/departure processes. When would burst-traffic patterns (market crashes, quest triggers) require a more sophisticated model?
What are the performance implications of possibility-map forking at scale — is there a practical limit on concurrent NPC possibility maps before memory becomes a constraint?