Editing Openai/693b7cce-38f4-800d-92f7-5e56a467bfa7 (section)

=== 6. Machine Learning Training Pipelines ===

You can train NPC models in three ways:

==== A. LLM-Driven NPC Dialogue & Decisions ====

Use a small model (3B–13B) and fine-tune it using:
* NPC internal state
* Recent world events
* Desired action + dialogue

Each training example looks like:

<syntaxhighlight>Input: NPC personality, emotions, beliefs, CK_depth, trust matrix, narrative role
Output: Action + Dialogue consistent with state

</syntaxhighlight>

This produces consistent personalities and adaptive dialogue.

==== B. Neural Surrogate Model (State Transition Network) ====

Train a tiny network that predicts:

<syntaxhighlight>state(t)  →  state(t+1)

</syntaxhighlight>

It learns the ODE/ABM dynamics and runs lightning-fast inside a game.

==== C. Reinforcement Learning NPC Policies ====

Train agents in a social-knowledge environment (“NPC Gym”):
* Rewards for narrative coherence
* Cooperation
* Emotional believability
* Survival / quest progression

This gives emergent behavior, alliances, grudges, betrayals, panic cascades, etc.