In 2D illustration an artist makes a series of decisions throughout the process, with those choices becoming more granular over time. These decisions vary in scope and impact. Given this formulation of a creative process, I observe that there exist portions of common workflows where an artist is largely following through on the logical conclusions of prior decisions. These are the specific areas of these creative workflows where I believe that co-creative tools can be implemented while retaining creative control over the completed artifact. My work so far has been in implementing such tools for steps in illustration and frame-by-frame animation. I intend to design further process-centered co-creative systems and study their usability and impact in the context of existing workflows.
Narrative planning can be used to create structured interactive experiences that dynamically respond to user input. Narrative planning works by generating a sequence of actions that achieves an author's desired goal while ensuring that there is an explanation for why each character takes each action in the sequence. An action in a sequence is considered necessary to that sequence if leaving the action out would prevent a later action in the sequence from being taken or prevent an author or character goal from being achieved. Using this definition, we define the causal width of a sequence to be the number of causally unnecessary actions, and we hypothesize sequences with a lower causal width are more likely to lead to a solution. We show that using causal width as a ranking mechanism can sometimes improve blind search, and ignoring stories with a high causal width can always improve the performance of heuristic search on a set of story benchmark problems.
We formalize the sensemaking needed to play information storygames as a task grounded in automated plan recognition, and outline extensions to a baseline model thereof needed to account for a key part of the information storygame-play loop: non-linear discovery of ambiguous information. This novel problem setting is non-trivial—mechanically simulating this narrative sensemaking requires piecing together incompletely-specified events to reason about the means through which particular ends were achieved in the virtual world. Our work readies plan recognition systems for the task by extending a foundational compilation of plan recognition as planning to cover partially-ordered and lifted observations of both actions and facts. While state-of-the-art plan recognizers approximate piecemeal aspects of these features, they do so in isolation of each other and by appealing to disparate and complex frameworks. In contrast, we achieve these functions within a classical planning-based framework. Our results confirm that, while slower, our approach never has more (and often has fewer) false positives than the baseline model in predicting the ground truth plan being executed. We discuss our findings in the context of future work toward better simulating human information storygame sensemaking.
Many recent advances in artificial intelligence remain difficult to apply in real games due to their black-box nature. These systems often lack transparency and control, making them difficult to integrate into games where player experience and narrative coherence are important. Unpredictable or inexplicable agent behavior can confuse players and frustrate developers. My dissertation explores how program synthesis can address this issue by generating interpretable, controllable representations of agent behavior. Instead of relying on black-box neural network policies, symbolic programs are extracted or generated that capture agent logic in a readable and editable form. Several methods are explored: synthesizing functional programs to imitate and explain game agents, adapting logical program policies to multi-agent settings, and evaluating large language models for code generation across domains such as simplified Atari games, Baba is You, and also tabletop games. An open research question is whether the created conceptual program libraries are transferable across different game domains. My research aims to bridge the gap between AI capabilities and game development needs by making agent behavior transparent, explainable, and adaptable for developers.
Narrative generation frameworks often face a trade-off between character believability and storyline adherence. Planner-based approaches ensure adherence to authorial designs at the cost of character agency and believability. In contrast, emergent narratives excel at presenting believable characters but often lack meaningful plot progression. We present DiriGent, a novel cognitive framework for agent modeling and narrative generation that enables authentic behavior while maintaining storyline adherence. Our agents possess a dynamic belief system and role-based ideal worlds that encode their basic values and relationships. We leverage Large Language Models (LLMs) to analyze the tensions between an agent's ideal worlds and their perceived actual world, which motivates their belief-driven actions. The system then adjusts the narrative world to amplify these tensions, thereby steering agent behavior toward the desired storyline. This dynamic framework allows agents to evolve meaningfully as the story unfolds, overcoming limitations of static agent profiles such as OCEAN. We evaluate our approach by generating stories for five story prompts. Our evaluation, consisting of automated LLM judges and human assessments, demonstrated significant improvements in character development and character motivation compared to baselines, while preserving storyline adherence. Our work presents a path toward interactive narratives that deliver rich characters and enable unique user experiences while adhering to desired storylines.
Developing game mechanics is challenging due to the need for intricate design and programming. Procedural Content Generation (PCG) is a prevalent aspect of modern video game development, enabling the generation of content via algorithms. Achieving the desired balance and player experience is a multifaceted challenge, with game mechanics playing a crucial role—requiring thorough testing, player feedback, and iterative refinement. This work explores automated approaches to mechanic generation and evaluation, drawing from Automated Game Design (AGD). I present methods for generating mechanics, reconstructing levels through level inpainting, and creating enemies that can only be defeated using newly generated mechanics. Comparative studies between reinforcement learning agents and traditional static agents such as A* show that RL facilitates more diverse and human-like mechanic discovery, while static methods remain more stable but less creative. Ongoing work integrates these techniques into environments where mechanics, levels, and enemies co-evolve, enabling richer evaluation of gameplay dynamics. To assess alignment between generated content and designer intent, I propose Design Impact Accuracy (DIA) as a metric to measure how effectively new mechanics are supported within AI-generated levels and enemies.
Game creation tools that minimize required resources and knowledge to use them have transformed the practice of learning and prototyping game design, especially in hobbyist and indie development contexts. However, little is known about how the different programming models found underlying these tools affect their expressiveness and usability. A recently proposed programming model allows creators to author the logic of the entire game using a single "game behaviour tree" with tile-grid rewrite rules at the leaves. We contribute to this body of knowledge by studying this recently proposed programming model. We have used it to make clones of popular games as case studies, from which we extracted a number of design patterns. To gain formative information about usability, we also conducted a small user study with people who have varying levels of experience authoring games with other tools. We find game behaviour trees capable of expressing a wide variety of 2D turn-based games. Study participants are quick to grasp the underlying concepts, but further research is needed to understand discrepancies with user intuitions that may arise from their familiarity with different programming models.
Generative design - the use of randomness, simulation or complex algorithms as part of a creative work - remains an uncommon skillset in the games industry. Acquiring familiarity and confidence with generative concepts is difficult and intimidating, which hampers its uptake in games. In this paper we argue that game design is generative design, and that by embracing the two as one and the same we can find new ways to communicate generative thinking to those who are unfamiliar with generative systems. We discuss the history of procedural generation in research and design, as well as reporting on surveys of 261 players and 126 game designers; we introduce the term procedural gameplay system to define a subset of generative systems used in games; and we offer some insights into how drawing an equivalence between these two design practices can yield new approaches and perspectives.
Interactive story systems today sit at three extremes. Emergent multi‑agent simulations give each character local intelligence but no global view, often losing plot structure. Reactive systems makes fast, state‑based decisions. They form plans using hand-authored rules without searching for action sequences, so these systems can respond quickly but can wander if long-term rules are not explicitly authored. Centralized narrative planners reason globally to craft coherent, goal‑directed plots, yet are computationally expensive. In my doctoral work I treat these not as isolated choices but as the three corners of a triangle spectrum of narrative generation. I propose hybrid, landmark‑guided approaches that can scale to larger domains. I am also exploring how large language models (LLMs) can be embedded within these hybrid approaches themselves. This paper outlines research questions, methodology, progress to date, evaluation plan, and requested feedback.
This paper presents a software library that enumerates the space of a state transition system specified by an action language, visualizes the states and action connections as a graph, and modifies the visualization based on underlying features determined through state and graph analysis. The library is intended as a tool for strong story interactive narrative design.
Most narrative planners currently lack the ability to model characters believing in multiple different possibilities or actions with different outcomes. By introducing a sense of uncertainty into narrative planning, we can model more realistic human behavior as well as generate new types of stories. In particular, stories surrounding a detective or a mystery require some form of uncertainty for the story to sound reasonable to a reader.
In this paper, I present a preliminary approach for encoding location data for CiF-like social characters in a to alleviate some of the authorial burden associated with authoring these types of interactive characters. To ease some of the authorial burden, I discuss how decoupling character actions and storing them within the environment can lead to a lowering of authorial complexity as well as treating the environment as a first-class entity can lead to more nuanced character performances. To better understand what is being decoupled, I describe how the term environment may be too vague when regarding interactive characters and their performances, and instead extending concepts from theater and dramaturgy, such as stage and props. From there I describe their functionality and possible encoding for props and stages.
In strong story experience management problems, an automated storytelling agent balances player autonomy with narrative structure in the context of an interactive story game world. However, it is possible for the game world to get softlocked in states outside narrative structures specified by the game designer. These states are called dead-ends. In this paper, we revisit adversarial strong story experience management, a framing of the experience management problem that models interactive storytelling as an adversarial game where dead-ends are losses. This framing is adversarial against narrative softlocks, not necessarily the player. We present a novel agent based on adversarial search and deep reinforcement learning, which is trained to avoid dead-ends while preserving player autonomy. We compare our approach to a reactive, narrative plan-based mediation system on a test set of games compatible with current narrative planning techniques. We show that our adversarial architecture outperforms narrative mediation on a suite of dead-end metrics during game trace and breadth-first tests of state transition system exploration, using classical and intentional planning domains.
Interactive storytelling offers personalized and engaging narrative experiences but poses significant authoring challenges. Our proposed framework, CoDi, extends the existing director-actor paradigm by enhancing the director agent's control capabilities. Specifically, CoDi uses high-level narrative goals to facilitate adaptive storytelling, with the director agent able to introduce new events, select relevant non-player characters (NPCs), and explicitly describe narrative outcomes. Comparative evaluations have demonstrated CoDi's competitive narrative quality, highlighting its effectiveness in balancing structural control and flexible agent behaviors, thus underscoring its potential as a foundation for scalable interactive storytelling systems. The code is publicly available on Github.
Struggle is an inevitable part of gameplay, and it’s often what makes games meaningful, rewarding, and fun. Still, some moments of difficulty spiral into frustration or confusion, and can cause players to quit entirely. Being able to detect and interpret struggle is thus essential for designing better player experiences, but identifying these moments remains challenging. We present a machine learning approach for detecting player struggle in real time using gameplay telemetry. Using three quests built in Terraria that each emphasize a different set of game mechanics – gathering, combat, or crafting – we collected data on how players interact with different systems and had them annotate where they encountered difficulty. Using this dataset, we trained Random Forest classifiers and evaluated model performance across different feature sets, window sizes, and step sizes. Our results show that such a model can successfully identify whether unseen players are experiencing struggle in the crafting quest, while the other quests proved more difficult. We also tested whether a model could classify the type of struggle as cognitive or performative, and found promising results for the crafting and combat quests. Our findings demonstrate the potential of using player telemetry to detect struggle, laying the groundwork for future adaptive systems that offer real-time, context-aware support tailored to individual player needs.
Narrative planning is the process of generating sequences of actions that form coherent and goal-oriented narratives. Classical implementations of narrative planning rely on heuristic search techniques to offer structured story generation, but often struggle with scalability because of large branching factors and deep search requirements. To improve the speed of narrative planning, we introduce Fog of War pruning, where Actions are only allowed if they involve people, places, and things that the protagonist character has discovered. This pruning technique restricts the planning to what is known from the perspective of the story's central character or characters, pruning branches of the search tree that involve actions beyond their current knowledge. This method is particularly useful in narratives where there is a strong protagonist focus and the story unfolds gradually as the character learns. This enables more efficient planning, while more closely aligning with how people would experience stories. Experiments across many narrative domains show that this technique not only speed up the search process, under identical search limits, also lets the planner solve more unique problems.
There has been an ongoing trend of pitching neural networks as playable game engines. However, from a game developer’s perspective, world models are not very practical since they cannot make changes like in a traditional game engine. Our project learns a model of pacman from gameplay video that can be played in pygame. The learnt model captures the original behaviour of the entities in the majority of the instances.
When generating levels for tile- and turn-based 2D games, generating a solution along with the level can ensure that the level is completable. One approach is to represent the level and solution as a 3D spacetime block with two spatial dimensions and one time dimension. Recent work in Space-Time WaveFunctionCollapse (STWFC) explored learning to generate such levels with solutions from example blocks by adapting 3D WaveFunctionCollapse for two spatial and one temporal dimension. While this generated solvable levels, the approach was found to be slow and could violate global constraints such as having a single player. Thus, in this work we explore learning to spacetime generate levels with solutions from examples by applying an existing 2D constraint-based level generation system. Rather than extending the 2D generator to 3D, we encode the time dimension in 2D by using a filmstrip representation where each frame represents a step in the time dimension. We compare STWFC and two filmstrip-based approaches: block constraints across frames that are analogous to those of STWFC, and constraints based on differences between frames. We demonstrate an application for interactive spacetime editing of levels and their solutions in the filmstrip representation, along with generating variations of levels sharing the same solution path, and variations of levels whose solution goes through the same arrangement of tiles at a point in time.
This paper outlines a supervised learning approach to model the combat outcomes of 1 vs 1 unit match-ups in strategy games. This is a core system in Creative Assembly's Total War. It is used to calculate combat outcomes in the auto-resolver and informs decisions of the battle AI whether units should engage or retreat from combat. We propose an alternative to the heuristic rules-based approach by using an automated framework to learn the combat outcomes. We describe our experiments, challenges and demonstrate that our model can accurately predict the unit's combat outcomes.
Simulating narrative comprehension has received little attention despite potentially aiding greatly in narrative generation or aiding in designing narratives. The work must be tied to theories of mind and understanding, and be based on algorithms that are described by psychology itself. For my thesis work, I propose building a cognition-like system based on the unified narrative understanding model: Multiversionality. In order to test the algorithmic model, I present the methods by which I plan to verify the solution: 1. by testing whether updates to prior knowledge that contradict prior assumptions actually affect, slow down, comprehension, 2. by testing how knowledgeable a player is in a given genre speeds up or slows down understanding by varying the input models, and 3. by testing how different refinement strategies could encapsulate human understanding.
