Room 1

This talk delves into how language models transform story narratives into fully realized game worlds. Starting with a story, we generate an entire game world including detailed maps, locations, and objects that align with the narrative. Furthermore, the world is populated with dynamic characters that have evolving memories, relationships, and behaviors shaped by the narrative. This approach can potentially complement existing generation tools to leverage the strength of different techniques.

The talk will feature a Murder Mystery demo, demonstrating an end-to-end generation pipeline going from story to a playable game, resulting in a meaningful generated world that reflects the intricate relationship between story, environment, and character evolution.

Takeaways:

• Using LLMs for narrative-driven generation for game development
• Combine LLMs with game mechanics to drive dynamic NPC behaviors
• Use cases where LLMs are most effective as a generation tool, along with practical tips and strategies

This talk is about Decima’s implementation of HTN. We will give a short history of our implementation and talk about how it integrates with the rest of our AI system. We compare it to behaviour trees and some other HTN implementations. We explain how our implementation performs backtracking (similar to Prolog) over preconditions and present a flow visualization which can help understand the backtracking flow, This talk will detail how that flow is realized in generated C++ and also touch on how we debug our HTN decompositions in-game.

Takeaways:

• Attendees will learn how HTN with backtracking can be realized in (generated) C++ code.
• Learn about a specific flow visualization (inspired by the “Byrd Box model” for Prolog) and how it can help understand backtracking flow.
• Learn how an HTN decomposition can be logged and how these logs can be displayed.

Description: Avalon is a new method designed to enhance level generation by providing more control over the generation process, while ensuring the creation of more solvable levels. In particular, we generate layouts of match-3 levels where the level designers can select visual features like size and symmetry and gameplay statistics such as difficulty. We will describe how we designed the system, how it compares to other methods and the quantitative and qualitative analysis performed on it. We will finalize with a brief discussion about challenges and opportunities of level generators in production.

Takeaways:

• How to create controllable (conditional) generators for match3 games
• How to take advantage of that controllability to improve the validity of the levels
• Challenges and opportunities of level generators in production

Description: A continuation and recontextualization of my 2024 GDC AI Summit talk “How ‘Ara: History Untold’ Transformed AI in 4X Games”, I will summarize Ara: History Untold’s novel system, and dig into details around optimizing its performance, and usability in the final months leading up to ship.

Takeaways:

• An introduction to semi-obscure but very useful AI topics (HGNs, Logical DSLs)
• Insights into making performant data structures for complex game AI
• Practical examples of how to make trades between effectiveness and complexity