Honors Capstone Project Report

Honors Capstone Prospectus

Kaveen Herath Bandara

Mentors: Kalpesh Padia, Dr. Christopher Healey

Tentative Capstone Title

Integrating Hierarchical Task Network Planning with Storyline Visualization

Tentative Final Product

SNCURCS 2017 poster presentation and a paper ready for EuroVis 2018 submission.

Problem

In recent years, researchers have used Artificial Intelligence to develop novel ways of generating narratives. Researchers in the AI domain have used Hierarchical task networks (HTNs), amongst other techniques, to generate interactive and engaging narratives. This project explores real time narrative visualization by integrating a HTN planner with YARN [1], a web-based visualization tool for evolving

Numerous systems for visualization of movie narratives and social stories have been created such as LifeLines [3], PlanningLines [4], EventRiver [5], TM method [6] and StoryFlow [2]. LifeLines and EventRiver are visualization systems for personal histories and text collections respectively. TM method is a set of design considerations along with a layout algorithm for producing visually pleasing storyline visualizations. Also, each of the above techniques represents narrative as node-link graphs, with nodes representing events, and links representing character movement. StoryFlow is an optimization technique to create fast and aesthetically pleasing storyline visualizations. The visual appeal or the legibility of a storyline visualization is measured by the number of lines crossings, line wiggles and efficient white space use. StoryFlow uses both discrete as well as continuous optimization to reduce the above mentioned metrics to improve legibility. Unlike previous visualization approaches, StoryFlow is capable of creating a visualization within seconds. For a story with hundreds of time frames and entities StoryFlow creates a visualization in a few seconds [2]. The StoryFlow algorithm is broken down into four distinct stages, where the first stage creates an initial relationship tree for time frames by taking into account the locations and characters at that specific time frame.

An HTN takes in a hierarchy of tasks and then decomposes each task in the hierarchy, starting with the initial task, until a solution is found by the planner. The domain knowledge required for the planning is structured as an AND-OR graph. HTN uses two different kinds of tasks for planning: primitive tasks and compound tasks. It takes in a list of operators, methods, and a list of tasks for the planning. An operator contains a primitive task along with one or more preconditions. Execution of an operator moves the planner from one state to the next. Compound tasks are a sequence of primitive tasks, and methods contain compound tasks along with a set of pre-conditions. Tasks that meet all of the preconditions are decomposed and added to the final solution. At the end of the planning the solution returned by planner is a list of tasks that must be executed to achieve the final