IC-951The programmatic space achieves behavior-similarity values comparable to the LEAPS latent space, indicating that optimizing the behavior loss alone does not produce a more search-conducive space

Tales Henrique Carvalho, Kenneth Tjhia, Levi Lelis

SourceReclaiming the Source of Programmatic Policies: Programmatic versus Latent Spaces

The authors measure the ρ-similarity (normalized longest common prefix of action sequences) between a program and its neighbor after 1–10 mutations, in both the programmatic space and the LEAPS latent space (σ = 0.1, 0.25, 0.5). They also measure the identity rate (probability a neighbor is the same program). The programmatic space achieves comparable behavior-similarity to the latent space across all σ settings, and the latent space with σ = 0.1, while achieving high similarity, has a higher identity rate. This shows that the behavior loss, which is the training objective of the LEAPS VAE, does not by itself guarantee a space that is easier to search.

Evidence

correlational

Caveat

The behavior-similarity metric is measured on 32 random Karel maps unrelated to any task and 1,000 initial programs; it captures local neighborhood structure but does not directly predict search performance, as the authors note: 'the observed result still does not explain the performance discrepancy that we observe when searching for policies that solve tasks.'

Model

LEAPS

Related work

LEAPS [compared-to]

Extraction

automatic-extraction