magistrsko delo
Abstract
Eden izmed načinov za modeliranje skupinskega vedenja je učenje s pomočjo genetskih algoritmov. Z njihovo uporabo lahko pridobimo modele, ki so uporabni za praktične namene.
V naši nalogi smo raziskovali učinkovitost genetskega programiranja za učenje vedenja pastirskih psov. V ta namen smo razvili simulacijo v okolju Unity, kjer je cilj psov, da čredo ovc preženejo v ogrado. Ovce so se obnašale po enem izmed dveh različnih režimov, ki smo jih implementirali, medtem ko so se psi vedli v skladu z drevesnimi strukturami, ki so bile naučene z genetskim programiranjem. Zanimalo nas je, kako različni parametri simulacije vplivajo na uspešnost modela. Nekateri izmed parametrov, ki smo jih spreminjali, so velikost populacije, prisotnost ovir v prostoru, različno obnašanje ovc ter spreminjanje lokacije ograde v prostoru. Ugotovili smo, da se najboljšega vedenja model nauči v preprostih pogojih. Opazili smo tudi, da je takšno vedenje učinkovito tudi, ko v prostor dodamo ovire. V primeru, da smo modele učili v bolj zahtevnem prostoru, smo ugotovili, da učenju pomaga, če psom omejimo vid.
Keywords
skupinsko vedenje;črede ovc;modeliranje skupinskega vedenja;simulacija;računalništvo in informatika;magisteriji;
Data
Language: |
Slovenian |
Year of publishing: |
2021 |
Typology: |
2.09 - Master's Thesis |
Organization: |
UL FRI - Faculty of Computer and Information Science |
Publisher: |
[Ž. Kotnik Klovar] |
UDC: |
004.414.23:636.7.043.7(043.2) |
COBISS: |
87405827
|
Views: |
212 |
Downloads: |
37 |
Average score: |
0 (0 votes) |
Metadata: |
|
Other data
Secondary language: |
English |
Secondary title: |
Developing behaviour of herding dogs using genetic algorithms |
Secondary abstract: |
One of the ways to model collective behaviour is to develop herding behaviour using genetic algorithms. With their use we can develop models which can be used for practical purposes.
In our work we researched how efficient is genetic programming for train sheepdog. For this purpose we developed simulation in Unity engine where sheepdog's goal was to herd sheep into a pen. Sheep were following one of two models of behaviour we developed while sheepdogs followed instructions given by tree structures that were developed with genetic programming. We were interested how different parameters impacted efficiency of trained model. Some of the parameters we used were size of population, presence of obstacles in the environment, model of behaviour of sheep and location of sheep pen. With our experiment we concluded that best models were trained on simple environments. We noticed that such models were still successful when we introduced obstacles into the simulation. In case we trained models in a more complex environment models were more successful if we limited sheepdog's vision range. |
Secondary keywords: |
genetic algorithm;genetic programming;collective behaviour;sheep herds;computer science;computer and information science;master's degree;Pastirski psi;Genetski algoritmi;Genetsko programiranje (računalništvo);Računalništvo;Univerzitetna in visokošolska dela; |
Type (COBISS): |
Master's thesis/paper |
Study programme: |
1000471 |
Thesis comment: |
Univ. v Ljubljani, Fak. za računalništvo in informatiko |
Pages: |
48 str. |
ID: |
13943495 |