|
|
Using reinforcement learning to learn how to play text-based games
Zelinka, Mikuláš ; Kadlec, Rudolf (advisor) ; Lisý, Viliam (referee)
The ability to learn optimal control policies in systems where action space is defined by sentences in natural language would allow many interesting real-world applications such as automatic optimisation of dialogue systems. Text-based games with multiple endings and rewards are a promising platform for this task, since their feedback allows us to employ reinforcement learning techniques to jointly learn text representations and control policies. We present a general text game playing agent, testing its generalisation and transfer learning performance and showing its ability to play multiple games at once. We also present pyfiction, an open-source library for universal access to different text games that could, together with our agent that implements its interface, serve as a baseline for future research.
|
|
|
Using Neural Networks to Determine Semantic Similarity of Two Sentences
Hrinčár, Peter ; Kadlec, Rudolf (advisor) ; Helcl, Jindřich (referee)
Figuring out the degree of semantic similarity between two sentences is important for many practical applications of natural language processing. The goal is to determine the similarity of sentences on a scale from "sentences are unrelated" to "sentences are equivalent". In this thesis we examined application of di erent neural network architectures to solve this problem. We proposed models based on Recurrent neural networks, which convert text sequence to constant sized vector. We followed up with suitable representation of unknown words. Our experiments showed that simple architectures achieved better results on the used dataset. We see a future extension of this thesis by using bigger training dataset. 1
|
|
|
DyBaNeM: Bayesian Model of Episodic Memory
Kadlec, Rudolf ; Brom, Cyril (advisor) ; Lim, Mei Yii (referee) ; Pilát, Martin (referee)
Title: DyBaNeM: Bayesian Model of Episodic Memory Author: Mgr. Rudolf Kadlec E-mail: rudolf.kadlec@gmail.com Department: Department of Software and Computer Science Education Supervisor: Mgr. Cyril Brom, Ph.D. Department of Software and Computer Science Education Abstract: Artificial agents endowed with episodic (or autobiographic) memory systems have the abilities to remember and recall what happened to them in the past. The existing Episodic Memory (EM) models work as mere data-logs with indexes: they enable record, retrieval and delete operations, but rarely organize events in a hierarchical fashion, let alone abstract automatically detailed streams of "what has just happened" to a "gist of the episode." Consequently, the most interest- ing features of human EM, reconstructive memory retrieval, emergence of false memory phenomena, gradual forgetting and predicting surprising situations are out of their reach. In this work we introduce a computational framework for episodic memory modeling called DyBaNeM. DyBaNeM connects episodic mem- ory abilities and activity recognition algorithms and unites these two computer science themes in one framework. This framework can be conceived as a general architecture of episodic memory systems, it capitalizes on Bayesian statistics and, from the psychological...
|
|
|
Activity recognition in a smart home setting
Fiklík, Vladimír ; Kadlec, Rudolf (advisor) ; Brom, Cyril (referee)
The aim of this work was to implement and compare several activity recognition algorithms which could be used in a smart home environment and would be able to determine the current activity of an observed subject (virtual agent) in the smart home using only data gathered by elementary observations of the environment. Such algorithms are useful in several areas, for example to improve behavior of various virtual agents, making them more aware of actions of the other agents. The algorithms used in this thesis are based on Dynamic Bayesian Networks and have ability to determine whether the observed activity has been completed or just interrupted. An easily extensible 3D interactive simulator of a smart home environment was created to meet the needs of activity recognition and used to gather data for the learning and testing phases of the algorithms. The test subjects were human-controlled virtual agents.
|
|
|
Advanced use of ACT-R in Pogamut
Zemčák, Lukáš ; Kadlec, Rudolf (advisor) ; Děchtěrenko, Filip (referee)
The requirements for virtual agents are more and more demanding. In order to manage the complex behavior of the agent, it's possible to take advantage of cognitive architectures which arised on the field neuroscience and artificial intelligence. This work examines PoJACTR library which links Pogamut library for developing intelligent agents in Unreal Tournament 2004 and jACT-R library which is Java implementation of one of the leading cognitive architectures ACTR. This work also studies certain agent implementation problems in PoJACTR and proposes a solution for them in form of debugging tools, which were subsequently implemented on an Eclipse IDE platform. In addition, it expands PoJACTR navigation and communication library modules for the game - Capture The Flag. As a validation, two agents (bots) were developed to play game, one in standard Pogamut and one in PoJACTR. When matched against each other in battle, Po- JACTR bot had comparable performance to a Pogamut bot. The results showed that debugging tools facilitated development process of PoJACTR agents.
|
|
|
User Friendly Envioronment for Dynamic Bayesian Networks
Vinárek, Jan ; Kadlec, Rudolf (advisor) ; Skřivánek, Zdeněk (referee)
Title: User Friendly Environment for Dynamic Bayesian Networks Author: Jan Vinárek Department: Department of Software and Computer Science Education Supervisor: Mgr. Rudolf Kadlec, Department of Software and Computer Science Education Abstract: For open source tools with the graphical interface which are focused on datamining and written in the Java language there is a small support for processing of sequential data. One of the most popular models used for processing of sequential data is the dynamic Bayesian network, with the use of its inference algorithms. The aim of the theoretical part of the thesis was to find a program which supports graphical interface for datamining with a simple control and library which imple- ments inference algorithms of dynamic Bayesian networks in the best way. The aim of the practical part was to design and to program the extension for the chosen program (RapidMiner) with the use of the found library (JSMILE). In the ex- tension the combination of uses of learning algorithm Expectation-Maximization and inference algorithm of dynamic Bayesian network was tested for prediction of sequential data. The combination was compared to the use of learning models Support Vector Machines and Decision Tree on two examples. Keywords: dynamic Bayesian network, sequential data, time series, Java
|
|
|
Adaptive Agent in a FPS Game
Witzany, Tomáš ; Kadlec, Rudolf (advisor) ; Hric, Jan (referee)
In this work I design and implement an adaptive oponent for the computer game Unreal Tournament for its Deathmatch mode. The agent has been designed using reinforcement learning and implemented on the Pogamut platform. A k-means clustering algorithm has been used for state abstraction. Furthermore an agent performance testing framework has been developed for the Pogamut platform aswell and used in this work. Several experiments testing different action-selection policies and different parameters of the Q-Learning algorithm were conducted. The resulting behaviour has a performance comparative to other implementations of reinforcement learning from other literature.
|
|
|
Advanced use of ACT-R in Pogamut
Zemčák, Lukáš ; Kadlec, Rudolf (advisor) ; Pešková, Klára (referee)
The requirements for virtual agents are more and more demanding. In order to manage the complex behavior of the agent, it's possible to take advantage of cognitive architectures which arised on the field neuroscience and artificial intelligence. This work examines PoJACTR library which links Pogamut library for developing intelligent agents in Unreal Tournament 2004 and jACT-R library which is Java implementation of one of the leading cognitive architectures ACTR. This work also studies certain agent implementation problems in PoJACTR and proposes a solution for them in form of debugging tools, which were subsequently implemented on an Eclipse IDE platform. In addition, it expands PoJACTR navigation and communication library modules for the game - Capture The Flag. As a validation, two agents (bots) were developed to play game, one in standard Pogamut and one in PoJACTR. When matched against each other in battle, PoJACTR bot had comparable performance to a Pogamut bot. The results showed that debugging tools facilitated development process of PoJACTR agents. Powered by TCPDF (www.tcpdf.org)
|
|
|
Coordination of multiple virtual agents in team-based computer games
Kolombo, Martin ; Gemrot, Jakub (advisor) ; Kadlec, Rudolf (referee)
Práce se zabývá problémem efektivní koordinace týmu virtuálních agent· v počí- tačových hrách. Hlavním cílem bylo navrhnou obecnou architektru pro práci s informacemi o virtuálním prostředí, která poskytuje programátorovi dobrý přístup k týmovým znalostem o virtuálním prostředí. Zvolená architektura neklade na programátora žádné požadavky ohledně implementace týmu agent· a nezávisí na žádném konkrétním virtuálním prostředí. Architektura je založena na přirozeném rozdělení dat podle jejich schopnosti se měnit v pr·běhu simulace a podle jejich subjektivity. Tato distribuce dat přirozeně vytváří sdílené znalosti týmu a programátorovi tak umožní založit rozhodování agent· na znalostech celého týmu namísto pouze jediného agenta. Architekturu jsme implemento- vali do systému Pogamut, ve kterém jsme provedli validaci zpětné kompati- bility. Vzheldem k nekompatibilitě formátu dat exportovaných součástí Poga- mutu, který jsme objevili v poslední fázi validace jsme provedli validaci pouze částečnou. Úpravu platformy Pogamut a následnou implementaci týmu agent· využívajícího naši architekturu plánujeme jako pokračování práce. 1
|
|
|
Episodic Memory with Believable Forgetting
Soukup, Tomáš ; Brom, Cyril (advisor) ; Kadlec, Rudolf (referee)
Title: Episodic Memory with Believable Forgetting Author: Tomáš Soukup Department: Department of Software and Computer Science Education Supervisor: Mgr. Cyril Brom, Ph.D., DSCSE Abstract: Presented thesis introduces a model of episodic memory for virtual humans extended by believable forgetting based on rating of memories according to their importance. It is inspired by a psychological model of E. Tulving and experiments of W.A. Wagenaar and M. Linton and builds on a memory model for a human-like agent developed by Klára Pešková, which we modified for the needs of forgetting. Our model takes advantage of the level-of-detail approach to forget the parts of memories gradually. In addition to the age and particularity of memories, emotiveness is also considered during the rating of memories. A simple emotional model was created for this purpose. The functionality of our model was verified by implementing it into a prototype application, which simulates the life of a virtual human in a virtual world. Our experiments showed that the behavior of our model, when configured properly, corresponds with the psychological concepts. Keywords: episodical memory, forgetting, virtual human, emotion, believability
|
guest ::
RSS feed.