Skip to yearly menu bar Skip to main content


Poster

Write, Execute, Assess: Program Synthesis with a REPL

Kevin Ellis · Maxwell Nye · Yewen Pu · Felix Sosa · Josh Tenenbaum · Armando Solar-Lezama

East Exhibition Hall B, C #170

Keywords: [ Applications ] [ Program Understanding and Generation ] [ Algorithms ] [ Structured Prediction ]


Abstract:

We present a neural program synthesis approach integrating components which write, execute, and assess code to navigate the search space of possible programs. We equip the search process with an interpreter or a read-eval-print-loop (REPL), which immediately executes partially written programs, exposing their semantics. The REPL addresses a basic challenge of program synthesis: tiny changes in syntax can lead to huge changes in semantics. We train a pair of models, a policy that proposes the new piece of code to write, and a value function that assesses the prospects of the code written so-far. At test time we can combine these models with a Sequential Monte Carlo algorithm. We apply our approach to two domains: synthesizing text editing programs and inferring 2D and 3D graphics programs.

Live content is unavailable. Log in and register to view live content