One-sentence Summary

This paper introduces a new modality for intelligent interactive theatre that employs an intelligent agent to serve simultaneously as a drama manager and a live actor, while identifying implementation challenges and potential methodologies to bridge gaps in current literature.

Key Contributions

• This work introduces an interactive narrative modality for live theatre wherein an intelligent agent simultaneously portrays a character and executes Drama Management tasks in response to human performers, positioning the audience as observers rather than active participants.
• The methodology integrates supervised learning to capture human acting patterns with generative models to produce diverse, archetype-consistent behaviors, and explores unsupervised Apprenticeship Learning to acquire broader narrative behavioral patterns.
• The approach extends traditional Drama Management systems to coordinate multiple live performers by evaluating Riedl et al.’s multi-subject management framework to identify avenues for reducing complexity while maintaining narrative coherence.

Introduction

Interactive narrative systems have long aimed to give users meaningful agency over story progression, yet traditional implementations like video games and choose-your-own-adventure formats require the audience to directly participate as protagonists. This participatory requirement restricts the medium's applicability to live performance, while existing drama management research primarily addresses single-player interactions and game-based non-player character believability rather than theatrical character archetypes. To bridge these gaps, the authors introduce a novel interactive theatre modality where an intelligent agent functions simultaneously as a drama manager and a live-performing character. The system responds to human actors in real time while guiding the narrative for an observing audience, leveraging supervised and apprenticeship learning to develop believable, archetype-aligned performances. This approach enables more dynamic and unpredictable storylines that sustain audience engagement and extend the lifespan of interactive experiences.

Method

The authors leverage a framework designed to integrate artificial intelligence into interactive narratives (INs), particularly within live theatre environments, by developing an AI actor capable of adapting to dynamic performance conditions while maintaining narrative coherence. Central to this approach is the use of a drama manager that mediates between player agency and authorial intent, ensuring the narrative experience remains engaging and consistent despite unpredictable player actions. The system builds upon the narrative structure of The Melete Effect, a choice-based narrative that serves as the foundational script for exploring multiple narrative permutations. This script is used to train the drama manager to anticipate and respond to a range of player decisions, enabling the system to guide the narrative along coherent and meaningful paths.

The proposed model incorporates player modelling techniques to capture and represent individual player preferences and behaviours, allowing the drama manager to tailor the narrative experience accordingly. This personalization is achieved through a character-specific representation that includes traits such as moral alignment, role in the story, and situational constraints, enabling the AI agent to behave consistently with its assigned character. These character models are integrated into the broader narrative framework, ensuring that responses remain contextually appropriate and aligned with the story’s thematic and structural constraints.

A key component of the system is its adaptability to unobserved or unplanned actions by performers. Unlike traditional text-based games, where unrecognized commands are simply rejected, live theatre presents a more complex environment where performers may deviate from scripted actions. To address this, the AI actor employs a multi-step strategy: first, it learns from a diverse set of training data to generalize across potential scenarios; second, it uses goal, plan, and action recognition to map novel events to known event types; and third, it relies on predefined behaviours for situations that fall outside its learned repertoire. This layered approach ensures that the narrative can evolve dynamically while preserving its overall integrity.

The architecture emphasizes the need for an AI actor that not only behaves in character but also adapts to real-time performance variations. This adaptability is essential for maintaining the immersive quality of the experience, particularly in live settings where disruptions are inevitable. By combining character modelling, player interaction analysis, and adaptive response mechanisms, the system aims to create a seamless integration of AI into live storytelling, enabling a new form of interactive narrative that responds to both player choices and unscripted events in real time.