Reproduce “Analogy making as amortised model construction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
David G. Nagy, Tingke Shen, Hanqi Zhou, Charley M. Wu, Peter Dayan
The authors propose a framework for constructing task-specific models by analogy, reusing relevant structure and partial solutions from earlier situations. In a Markov decision process, an analogy maps parts of a source model to a target model. Exact shared abstract structure can support optimal policy transfer, but partial analogies do not guarantee success. The framework is conceptual, not experimentally tested.
Key insights
- The 'on-demand objective' formalizes model construction itself as a resource-constrained optimization problem, balancing model quality with the costs of construction and solution.
- Reusing abstract modules through analogy provides a powerful inductive bias, reducing policy search space and immediately constraining actions in novel situations.
- Modules in the library undergo continuous refinement and abstraction, evolving from concrete experiences to more generalized conceptual primitives, akin to the 'career of a metaphor'.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.