Protagonist vs Antagonist PROVANT: Narrative Generation as Counter Planning

Our motivation in this work is to develop a narrative generation mechanism for Interactive Storytelling that removes some of the authoring burden that is inherent to plan-based approaches. We focus on the class of narratives that dominate in Hollywood movies, television serial dramas and situation comedies. These narratives revolve around a central Protagonist in pursuit of a goal and who faces a series of obstructions placed in their way by an Antagonist and which they must overcome in order to reach their goal. We cast this problem as a non-cooperative multi-agent planning problem, in other words counter planning. We build on recent techniques in goal recognition and landmark identification to develop a novel plan-based narrative generation mechanism. A key opportunity that goal recognition provides is to reason explicitly with partially observed action sequences, reflecting the reasoning process of the antagonist. Thus the antagonist can only act to obstruct if it is reasonable (to the viewer) that they have guessed the protagonist's intentions. Starting from the believed goal, the narrative generator can reason about the protagonist's plan and what must be done to achieve it i.e., the plan landmarks and use these to automatically identify suitable points of obstruction. In the paper we detail the approach and illustrate it with a worked example. We report the results of an experimental evaluation and user study in a number of representative narrative domains. Results of the user study with system generated narratives confirm that viewers can clearly recognise agent roles and narrative structure.