Bayesian Optimization Framework for Open-Ended Task Discovery
The Generate-Select-Refine (GSR) framework employs Bayesian optimization to identify and enhance tasks within scientific workflows dynamically. In contrast to conventional BO, which operates under the assumption of a static objective, GSR alternates between the creation of new tasks and the refinement of current ones, beginning with a user-supplied seed task. This approach asymptotically focuses evaluations on the optimal task, incurring only logarithmic regret compared to single-task BO. GSR surpasses existing LLM-based optimizers in various applications, such as new product development, scaling chemical synthesis, analyzing algorithms, and repurposing patents. This research has been made available on arXiv in the fields of computer science and artificial intelligence.
Key facts
- GSR stands for Generate-Select-Refine
- It is an open-ended Bayesian optimization framework
- Alternates between task generation and task optimization
- Starts from a user-provided seed task
- Generates new tasks in a coarse-to-fine manner
- Uses a task-acquisition function to schedule optimization
- Asymptotically concentrates evaluations on the best task
- Logarithmic regret overhead relative to single-task BO
- Applied to new product development, chemical synthesis scaling, algorithm analysis, and patent repurposing
- Outperforms existing LLM-based optimizers
- Published on arXiv under Computer Science > Artificial Intelligence
Entities
Institutions
- arXiv