In this thesis, the concept of reinforcement learning from human advice is applied to the task of model tuning. The overall objective is to integrate evaluative feedback into the training process of reinforcement learning to fine-tune a Multi-Layer-Perceptron model according to specific stakeholder requirements. The concept is especially effective for natural, unconstrained requirements, which are hard to define mathematically but easy to demonstrate and recognize by humans. The usability and feasibility of this approach is analyzed in the concrete use case of class-based localization in the automotive domain.
Klíčová slova
-
Klíčová slova v angličtině
-
Rozsah průvodní práce
129 p. thesis, 15 p. appendix
Jazyk
AN
Anotace
-
Anotace v angličtině
In this thesis, the concept of reinforcement learning from human advice is applied to the task of model tuning. The overall objective is to integrate evaluative feedback into the training process of reinforcement learning to fine-tune a Multi-Layer-Perceptron model according to specific stakeholder requirements. The concept is especially effective for natural, unconstrained requirements, which are hard to define mathematically but easy to demonstrate and recognize by humans. The usability and feasibility of this approach is analyzed in the concrete use case of class-based localization in the automotive domain.
Klíčová slova
-
Klíčová slova v angličtině
-
Zásady pro vypracování
-
Zásady pro vypracování
-
Seznam doporučené literatury
-
Seznam doporučené literatury
-
Přílohy volně vložené
-
Přílohy vázané v práci
ilustrace, grafy, schémata, tabulky
Převzato z knihovny
Ne
Plný text práce
Přílohy
Odůvodnění nezveřejnění VŠKP
Posudek(y) oponenta
Hodnocení vedoucího
Záznam průběhu obhajoby
second opponent 1
Presentation of the student
introduction of used methods
Vohnoutova - how the human preference influence the results and where are the limitations