izpis_h1_title_alt

Use of relaxed stochastic controls in reinforcement learning : magistrsko delo
ID Rems, Jan (Author), ID Agram, Nacira (Mentor) More about this mentor... This link opens in a new window, ID Košir, Tomaž (Comentor)

.pdfPDF - Presentation file, Download (794,72 KB)
MD5: F758F2DEA8454D2B99EEBBEA38C1DBB4

Abstract
In this work, we investigate how relaxed stochastic controls are used for exploration in continuous time and space reinforcement learning. The environment $X^u$ is modeled by a stochastic differential equation controlled by control $u$, while the value function $V^u$ is an infinite horizon performance functional. For relaxed control distribution $\pi$ we introduce relaxed versions of environment $X^{\pi}$ and value function $V^{\pi}.$ In a special linear-quadratic case the optimal control distribution turns out to be Gaussian with mean depending on the current state, and variance depending on exploration weight parameter. A reinforcement learning algorithm for optimal investment strategy in a simple model of the financial market with the infinite horizon is developed and tested.

Language:English
Keywords:reinforcement learning, exploration, stochastic control theory, relaxed controls, dynamical programming, optimal investment strategy
Work type:Master's thesis/paper
Typology:2.09 - Master's Thesis
Organization:FMF - Faculty of Mathematics and Physics
Year:2021
PID:20.500.12556/RUL-130550 This link opens in a new window
UDC:519.8
COBISS.SI-ID:79333891 This link opens in a new window
Publication date in RUL:16.09.2021
Views:1029
Downloads:196
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Uporaba relaksiranih stohastičnih akcij v spodbujevalnem učenju
Abstract:
V tem delu si ogledamo, kako uporabiti relaksirane stohastične akcije pri definiranju raziskovanja v spodbujevalnem učenju v zveznem prostoru in času. Prostor $X^u$ je modeliran s stohastično diferencialno enačbo kontrolirano z akcijo $u.$ Funkcijo vrednosti $V^u$ je funkcional uspešnosti na neskončnem časovnem obdobju. Za relaksirano akcijo $\pi$ vpeljemo raziskovalno verzijo okolja $X^{\pi}$ in funkcijo vrednosti $V^{\pi}.$ V posebnem linearno-kvadratičnem primeru se izkaže, da je optimalna relaksirana akcija Gaussova, kjer je pričakovana vrednost odvisna od trenutnega stanja, varianca pa od parametra, ki kontrolira raven raziskovanja v modelu. Predstavljen je algoritem spodbujevalnega učenja za napoved optimalne strategije v preprostem modelu finančnega trga z neskončim časovnim oknom.

Keywords:spodbujevalno učenje, raziskovanje okolja, teorija upravljanja stohastičnih sistemov, relaksirane stohastične akcije, dinamično programiranje, optimalna investicijska strategija

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back