Inverting the Observation Model: How to Generate Code from Any Output

Table of Links

Abstract and 1. Introduction

1. Background & Related Work

2. Method

   3.1 Sampling Small Mutations

   3.2 Policy

   3.3 Value Network & Search

   3.4 Architecture

3. Experiments

   4.1 Environments

   4.2 Baselines

   4.3 Ablations

4. Conclusion, Acknowledgments and Disclosure of Funding, and References

   \

Appendix

A. Mutation Algorithm

B. Context-Free Grammars

C. Sketch Simulation

D. Complexity Filtering

E. Tree Path Algorithm

F. Implementation Details

3.2 Policy

3.2.1 Forward Process

\

\ 3.2.2 Reverse Mutation Paths

\ Since we have access to the ground-truth mutations, we can generate targets to train a neural network by simply reversing the sampled trajectory through the forward process Markov-Chain, z0 → z1 → . . .. At first glance, this may seem a reasonable choice. However, training to simply invert the last mutation can potentially create a much noisier signal for the neural network.

\ Consider the case where, within a much larger syntax tree, a color was mutated as,

\

\

:::info Authors:

(1) Shreyas Kapur, University of California, Berkeley ([e-mail]);

(2) Erik Jenner, University of California, Berkeley ([e-mail]);

(3) Stuart Russell, University of California, Berkeley ([e-mail]).

:::

---

:::info This paper is available on arxiv under CC BY-SA 4.0 DEED license.

:::

\