Creating Emergent Gameplay with Autonomous Agents Borut Pfeifer

Disclaimers and other Miscellaneous Warnings Not an AI programming talk. . . Not entirely a game design talk, either. Me: – Radical, 2003+ – White Knuckle Games 2001 -3 – 2 articles in Game Programming Gems 4 (one on DDA)

What the hell is Emergent Gameplay? Not emergent behavior or AI. A large amount of gameplay experiences from a much smaller set of interconnected game rules. Examples. . .

Chess 6 pieces, almost infinite gameplay scenarios.

Grand Theft Auto Rules allow the player to explore the world and accomplish set gameplay goals in many ways. GTA: Vice City Rockstar North Rockstar Games

HALO: Combat Evolved Linear experience, but emergent gameplay allows for dynamic situations and replayability. HALO Bungie Microsoft

So what? Emergent gameplay has advantages (but not the only style of gameplay): – Less scripted play, more replayability. – More reuse of resources (no such thing as mission specific gameplay or assets). How can we build AI to enhance this style of gameplay? What are the design issues involved?

Perspectives/Schema of Agent Behavior: Agent Behavior as Opponent AI needs to be as smart as possible to beat the player. AI's main purpose is to provide challenge – player wins the game by beating the AI. – Hardcore FPS AI – Chess programs (agent behavior isn't what encourages emergent gameplay). Not as valid for emergent gameplay – player wins the game or has fun by other means.

Perspectives/Schema of Agent Behavior: Agent Behavior as Game Rule Game goals achieved by strategic application of rules. Not because they’re “stupid” (limited senses, lack of context sensitive behavior). – HALO: Elites are tough in melee combat, sneak around them. – GTA: Avoid cops as you try to achieve ancillary goal Emergent gameplay – once player learns one rule, he/she can apply the same rule to new contexts. Agent behavior needs to be learned by the player.

Perspectives/Schema of Agent Behavior: Agent Behavior as Interface Player manipulates AI to achieve game goals (like the controller interface, but at a higher level). Understanding user's conceptual model: – visibility, mappings, feedback – Affordances - tough guys look tough – Constraints/forcing functions – ex. force player to move by getting more accurate the longer they stand still.

Designing Agents For Emergent Gameplay Orthogonal behavior – Combined behaviors expands gameplay possibility space (chain reactions). – Allows for a variety of player strategies. Teaching Behavior – Player must understand agent behavior to learn gameplay rules to apply them to new situations. – Agent's Sense – Think – Do cycle relates to player's own Sense - Think – Do cycle. – Player needs know the causes of behaviors (reactive). – Behavior Archetypes - Groups of agents that share, and appear to share, behavior - helps player learn.

Sense – Sensory Modeling an agent’s senses -> game rules Problems with conflicts (different sense having different priorities under different circumstances). Often sacrifice realism for gameplay (stealth games). Biggest area to develop for emergent gameplay AI agents need to be able to handle more context (react to more things, and remember more things).

Think – Decision making The Usual Suspects State Machines – Easy to embody game rules as states & transitions. – Causes state errors – enemy doesn’t know how to react to a stimulus b/c someone forgot to make it a trigger in the state he’s in. Can be decreased with augmentations (hierarchical, parallel, stack based). Behavior/Task arbitration – Harder to embody game rules, deals with conflicting contexts much better. – Problems with priorities – tasks flipping back and forth, hard to debug

Think – Decision making Phoebe Sengers - “Schizophrenia and Narrative in Artificial Agents”, Narrative Intelligence Agent behavior can be disjoint with no focus on what it is communicating. Yet another layer in our agent? For example: – Rules based system relating sensory stimuli to responses – focuses on what needs to be communicated to the player – Traditional method - handles “behind the scenes” decision making, resolving conflicts between what the agent is doing and what it needs to be doing/communicating.

Do – Communicating Intent Must communicate behavior through: – – – Action (what they do) Movement (where they move, how they move, speed) Animation Sound (effects and dialog) Modeling/texturing What does the player think the agent is thinking? – How does the player link their actions to agent behavior? – Can also record what we’ve communicated to the player to help track what they’ve learned.

The Player's Mental Model Dealing with Causality – Mind tends to say link event A caused event B if A happens just before B. – Need restrict agent behavior to being affected only by direct stimulus/action (preferably the player's). What sort of strategies is the behavior enabling? – What actions does it reward or what is the best action to deal with it? – Increase opportunities for emergence by making strategies conflict or interact. – Example from HALO: Elites require stealth, Grunts are easier to just overpower – mix them in the same combat and player must make decision.

Non-Deterministic Decision Making (or, You're all going to hate me for saying this. . . ) Bad - breaks consistency required for the player to learn game rules for emergence. For one set of input stimuli, the game needs to react the same way. How can the player learn the game if something different happens each time they do one particular thing?

Non-Deterministic Decision Making Random decision making occasionally has uses: • Novelty – Humor – Variety • Appearance of Depth – Exploit player’s inability to reason about random events. – Fakes complexity (pedestrians randomly waving to each other makes a city seem more involved). – Shouldn't affect core game rules of agent behavior.

Summary Easy to expand a gameplay space with reactive agent behavior. (Car swerves to avoid player, hits obstacle, explodes, bystanders die, cops come). To handle reactivity, more complexity is needed in the decision making layer, focusing on what we need to communicate. The player has to be able to learn the behavior & it's motivations in order to use it as a rule or in a strategy.

References Katie Salen and Eric Zimmerman - Rules of Play: Game Design Fundamentals. Phoebe Sengers - "Schizophrenia and Narrative in Artificial Agents", Narrative Intelligence. Harvey Smith & Randy Smith – "Will the Real Emergent Gameplay Please Stand Up? ", GDC 2004. Harvey Smith - "Orthogonal Unit Design", GDC 2003. Harvey Smith - "Systemic Level Design", GDC Europe 2002. Chris Butcher and Jaime Griesemer - “HALO: AI & Level Design”. GDC 2002 Email me - borut_p@yahoo. com