L-MCTS Framework of Massive Spatio-temporally Plausible Content Generation

   Spatio-temporal content, cascades of physical and procedural events, is one of the most indispensable elements in games and animations nowadays. Generally, to design such scenes needs expertise to achieve both procedural rules and targeting criteria such as complexity, appealingness, and difficulty, and hence, to generate various instances is time-consuming and manpower intensive. Therefore, we propose a general framework, named L-MCTS, whose automatic workflow, integrating expertise and domain knowledge, is capable of massive production with targeted criteria. More specifically, in order for systematic exploration, we break the content space of near-infinity degrees of freedom into three parts: forming cascading events with L-system, positioning elements with inverse embedding, and activating the content with initial conditions. Furthermore, we adjust Monte Carlo Tree Search (MCTS) to efficiently, systematically, and coherently sample plausible instances through the decomposed space. In order to automatize the work flow with objective evaluations, we empirically describe a metric development procedure derived from experiences and questionnaire postulation to encode the targeted criteria using content-describing tree complexity and parameter variations. Finally, we demonstrate the generality of our framework with three different applications and evaluate their effectiveness and efficiency through several experiments and user studies.