Learning Action Value Functions for Accelerating Multi-agent Planning for Emergent Narrative

Abstract

Emergent narrative is an approach to interactive storytelling where the story emerges out of a simulation. This thesis builds on an autonomous agent AI engine that was developed for a previous project in emergent narrative. The engine is still limited by the combinatorial explosion in the planning, especially for the large state spaces that realistic game universes require. In this thesis we study the action value function, which assigns each action a value that should represent the relative importance of that action compared to other actions, in order to allow for better planning while using less resources. We also look at first steps towards automatically optimising the actions the agent considers while planning, which should bring us closer to making emergent narrative more scalable.