Planning with Learned Binarized Neural Networks Benchmarks for MaxSAT Evaluation 2021

TL;DR

This paper introduces a framework for automated planning using binarized neural networks, provides a MaxSAT encoding for such problems, and presents four benchmark domains for evaluation.

Contribution

It presents a novel MaxSAT encoding for planning problems with BNNs and introduces new benchmark domains for MaxSAT evaluation.

Findings

Four benchmark domains introduced: Navigation, Inventory Control, System Administrator, Cellda

MaxSAT encoding effectively models planning with BNNs

Framework facilitates evaluation of planning algorithms with neural network-based transitions

Abstract

This document provides a brief introduction to learned automated planning problem where the state transition function is in the form of a binarized neural network (BNN), presents a general MaxSAT encoding for this problem, and describes the four domains, namely: Navigation, Inventory Control, System Administrator and Cellda, that are submitted as benchmarks for MaxSAT Evaluation 2021.