Project page for Setting Switcher: Changing genre-settings in text-based game environments populated by generative agents as presented at NeurIPS 2023 Workshop on Machine Learning for Creativity and Design.
This work is still ongoing. If you'd like to work together on further developing this to deploy in-game, please be in touch - let's make it happen.
Contents:
- Testing environment
- Example notebook
- Agent (switcher.py)
- All relevant modules for agent (locator, plans, etc)
Points for future work:
Big TODO:
- This work, particularly simulation.py, has since been significantly improved with added capability for perception and interaction. Updates to be made soon.
Dreams of further development and ideas for future work:
-
Update since-improved simulation functions with perception and agent interaction.
-
Implement higher quality generative agents, such as Concordia agents.
-
Permit user interactions within the game environment
-
Improve quality and accuracy of output further, testing more genres and tuning heuristic prompts.
-
Connect to the visual using a pipeline of GPT-4-Vision, stable diffusion/midjourney/any text-to-image and SettingSwitcher agent.
-
Build in additional functions for generating prompts tailored to X model based on setting i.e art style.
-
Work this into a rudimentary visual game environment.
-
-
Connect to the creation of location trees through prompting
- Using the description of locations on our map, generate sub-locations and items.