Paper Image

Repairing underwater vehicle controllers

Published on:

6 November 2023

Primary Category:

Systems and Control

Paper Authors:

Pengyuan Lu,

Matthew Cleaveland,

Oleg Sokolsky,

Insup Lee,

Ivan Ruchkin

Bullets

Key Details

Proposes Repair with Preservation (RwP) problem formulation to preserve correct behaviors during neural network controller repair

Designs Incremental Simulated Annealing Repair (ISAR) algorithm to solve RwP problem

Protects verified successful regions while repairing others using simulated annealing

Preserves 100% of verified regions in case studies while repairing over 80% and 20% of failures

AI generated summary

Repairing underwater vehicle controllers

This paper proposes a technique to repair neural network controllers for underwater vehicles. It identifies an issue with prior repair methods, which may break previously correct behaviors when fixing errors. The authors formulate a new optimization problem, Repair with Preservation (RwP), which aims to preserve correct behaviors while repairing others. They design an algorithm, Incremental Simulated Annealing Repair (ISAR), to solve this problem. ISAR protects verified successful regions and repairs additional regions using simulated annealing. Case studies on an underwater vehicle and mountain car simulation show ISAR is able to preserve all verified regions while repairing over 80% and 20% of failed regions, respectively.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up