Causal Inductive Synthesis Corpus

Zenna Tavares, Ria Das, Elizabeth Weeks, Kate Lin, Joshua B. Tenenbaum, Armando Solar-Lezama

Keywords: Program induction, Causal model discovery, scientific discovery, benchmark

Abstract:

We introduce the Causal Inductive Synthesis Corpus (CISC) -- a manually constructed collection of interactive domains.
CISC domains abstract core causal concepts present in real world mechanisms and environments. We formulate two synthesis challenges of causal model discovery: the passive discovery of a model of a CISC domain from observed data, and active discovery while interacting with the domain. CISC problems are expressed in Autumn, a Turing-complete programming language for specifying causal probabilistic models. Autumn allows succinct expression for models that vary dynamically through time, respond to external input, have internal state and memory, exhibit probabilistic non-determinism, and have complex causal dependencies between variables.