Indoor Visual Navigation using Deep Reinforcement Learning: Deep Reinforcement Learning
Author
Skov, Søren
Term
4. semester
Education
Publication year
2018
Pages
95
Abstract
Dette projekt handler om at lære en agent at navigere i indendørs omgivelser ud fra det, den ser. Tilgangen er deep reinforcement learning, hvor en neuralt netværksbaseret model lærer ved forsøg og belønninger og trænes på billeder med målet om at finde flere bestemte positioner. Arbejdet bygger på en tidligere opsætning, hvor en agent blev trænet på 100 millioner billeder for at finde 100 mål. Resultaterne viser, at der stadig er plads til forbedringer. Derfor analyserer vi, hvad agenten faktisk lærer under træningen, for at forstå, hvordan dette påvirker navigationspræstationen. Som led i analysen foreslår vi en måde at visualisere, hvad agenten lærer over tid, for at gøre det lettere at identificere styrker og svagheder.
This project aims to teach an agent to navigate indoor environments based on what it sees. The approach uses deep reinforcement learning, where a neural network learns via trial and reward, trained on images to reach multiple target positions. The work builds on a prior setup in which an agent was trained on 100 million images to find 100 targets. The results indicate there is room for improvement. To address this, we analyze what the agent actually learns during training to understand how this affects navigation performance. As part of the analysis, we propose a method to visualize what the agent learns over time, making it easier to pinpoint strengths and weaknesses.
[This abstract was generated with the help of AI]
Documents