GDnD Wiki

GDnD Wiki

This site is for Ulster University students and staff. Enter the class access password to continue.

Game AI Pro 360: Guide to Tactics and Strategy

5 min read

Game AI Pro 360: Guide to Tactics and Strategy

Source metadata

  • Type: Academic/industry textbook anthology
  • Editor: Steve Rabin
  • Publisher: CRC Press / Taylor & Francis Group, 2020
  • ISBN: 978-0-367-15088-4
  • Companion site: http://www.gameaipro.com

An anthology drawn from the three Game AI Pro volumes, focused on combat-decision AI for RTS, RPG, MOBA, strategy, and tower-defence genres. Contributors are primarily industry practitioners and academic researchers.

Key takeaways

  • Tactical position selection is best structured as a query language with three phases: generation (candidate points), filtering (conditions that discard invalid points), and weighting (scoring remaining points). A domain-specific language on top of Lua or similar scripting makes queries readable and rapid to iterate.
  • Cover representation can be stored as line segments (cover segments) associated with NavMesh polygons via a CoverMap lookup — this avoids ray-cast overhead during pathfinding.
  • Tactical pathfinding on a NavMesh adds a cover-bias factor to A* segment costs proportional to the exposed path length, computed by clipping the path segment against a frustum built from cover segment and enemy position.
  • The “Belgian AI” / Kung-Fu Circle system uses a grid of slots around the player with grid capacity and attack capacity budgets to control how many enemies can simultaneously approach and attack. A centralised stage manager assigns slots, enabling natural flanking without per-NPC coordination code.
  • Hierarchical AI for Killzone 3 bots uses three layers — individual bot (HTN planner), squad (HTN planner using strategic graph + influence map), and commander (objective assignment) — communicating via orders down and completion messages up.
  • HTN planning (Hierarchical Task Network) is favoured for bot AI because it gives explicit control over plan priority ordering; the domain decomposes from a root behave task through branches for self-preservation, class abilities, and combat.
  • Modular tactical influence maps separate proximity maps (where can agents reach) from threat maps (what can they threaten) and use stamped templates to avoid repeated distance calculations; working maps assemble combinations of base maps for local queries.
  • Spatial reasoning for strategic AI uses region-based map partitioning. Regions support chokepoint/cul-de-sac detection, hierarchical pathfinding, and region-level influence propagation — giving RTS-style AI emergent flanking and weak-point targeting.
  • Infinite-resolution influence mapping replaces the discrete grid with point-based falloff functions queried via a k-d tree, giving arbitrary spatial resolution at linear memory cost in the number of sources.
  • Monte Carlo Tree Search (MCTS) struggles with large strategy game action spaces; the solution is hierarchical expansion — decomposing moves into actions and expanding in order of apparent importance — combined with heuristic-guided playouts and abstract simulation models.
  • Combat outcome prediction using Lanchester’s attrition laws (linear and square law) provides orders-of-magnitude faster prediction than simulation; unit strength parameters can be learned from game replays using logistic regression.
  • Hierarchical Portfolio Search (HPS) for Prismata combines a portfolio of partial-player functions (one per tactical phase) with a top-level MCTS or alpha-beta search, reducing the action space while remaining robust to game-balance patches.
  • NPC searching uses a two-phase approach: Phase 1 investigates the last known position; Phase 2 sweeps a coordinator-managed pool of cover-derived search spots, with passive LOS checks clearing spots as NPCs pass through them.
  • Perception modelling (Splinter Cell Blacklist) uses coffin-shaped detection volumes, bone-raycast visibility, tiered bark systems, and TEAS (Tactical Environment Awareness System) connectivity graphs for area-aware sound propagation.

Notable claims

“Tactical Position Selection is a keystone of shooter AI and a potential Swiss army knife of AI and gameplay programming.” (Jack, Chapter 1)

“Because the influence system gives the AI some warning that an attack is approaching — the enemy influence rises — we can avoid [the problem of defenders not being present]. When the enemy launches their massive attack, the AI will know that it’s coming.” (Dill, Chapter 10)

“MCTS has a poor track record for tactical decision-making. More accurately, it struggles with problems that have a very narrow path of victory or success.” (Roelofs, Chapter 16)

“It is only important what’s plausible from the player’s point of view — it really doesn’t matter what the NPC should see or hear; it’s what the player thinks the NPC can see and hear.” (Walsh, Chapter 7)

Relevance

This source directly informs the following wiki topics:

Open questions raised

  • Chapter 1 notes the stark trade-off between decentralised individual position selection and centralised squad-level selection. Is there a middle ground with minimal coupling?
  • Lanchester models (Chapter 13) assume stationary engagement; how do they generalise to highly mobile, kiting-based RTS combat?
  • Chapter 16 proposes entropy-based ordering for MCTS action sets — what is the overhead of computing this online vs. offline?

Graph View

Backlinks