Loading...
Search for:
deep-reinforcement-learning
0.006 seconds
Total 27 records
Algorithmic Trading Using Deep Reinforcement Learning
,
M.Sc. Thesis
Sharif University of Technology
;
Marvasti, Farohk
(Supervisor)
Abstract
Price movement prediction has always been one of the traders’ concerns in the field of financial market prediction. In order to increase the profit of the trades, the traders can process the historical data and predict the movement. The large size of the data and complex relations between them lead us to use algorithmic trading and artificial intelligence.The stock and Cryptocurrency markets are two common markets attracting traders. This thesis aims to offer an approach using Twin-Delayed DDPG (TD3) and daily close price in order to achieve a trading strategy. Unlike the previous studies using a discrete action space reinforcement learning algorithm, TD3 is a continuous one offering both...
An Application of Deep Reinforcement Learning for Ambulance Allocation to Emergency Departments under Overcrowding Situation
, M.Sc. Thesis Sharif University of Technology ; Akhavan Niaki, Taghi (Supervisor)
Abstract
In the last decade, emergency department (ED) overcrowding has become a national crisis for the US healthcare system. Increasing mortality rates, decreasing quality of care, financial losses due to walkouts, and ambulance diversion are some of the consequences of ED overcrowding. Given the increasing demand in terms of ambulance utilization which we can see an instance of it in the COVID-19 pandemic, being able to allocate service requests to EDs efficiently, becomes a key function of emergency medical services. in this investigation, an algorithm of deep reinforcement learning called deep Q-learning is used to address this problem and to assign ambulances to ED's appropriately. under...
Meta Reinforcement Learning for Domain Generalization
, M.Sc. Thesis Sharif University of Technology ; Rohban, Mohammad Hossein (Supervisor)
Abstract
Deep reinforcement learning has achieved better cumulative rewards than humans in many environments like Atari. One drawback of these methods is their data inefficiency which makes training time-consuming, and in some cases having this amount of data is infeasible. Meta reinforcement learning can use past experiences to enable agents to adapt to new tasks faster and makes neural networks to train in a short amount of time.One of the methods in meta reinforcement learning is inferring tasks which helps exploitation policy to have good performance in new tasks. There’s a need to improve exploration policy as well as exploitation policy by gaining informative transitions about the new task....
Distributed Cache Management Using Reinforcement Learning based Strategies
, M.Sc. Thesis Sharif University of Technology ; Mir Mohseni, Mahtab (Supervisor) ; Maddah Ali, Mohammad Ali (Supervisor)
Abstract
Nowadays, video on demand causes a drastic increase in network traffic that it is expected that network traffic surpasses 45 exabytes per month until 2022; consequently, utilizing distributed memories known as caches across the network to alleviate the communication load during peak hours is inevitable. Coded caching is a promising approach to mitigate and smooth traffic during peak hours in the communication network in a way that it creates coded multicasting opportunities in addition to delivering content to users locally. However, it suffers from imposed delay resulting from coding that makes this approach infeasible for delay-sensitive contents, namely video streaming applications. So...
Optimal Process Planning for Automated Robotic Assembly of Mechanical Assembles based on Reinforcement Learning Method
, M.Sc. Thesis Sharif University of Technology ; Khodaygan, Saeed (Supervisor)
Abstract
Nowadays, the assembly process is planned by an expert and requires knowledge and it is time-consuming. The flexibility and optimality of the assembly plan depend on the knowledge and creativity of the expert, and therefore expertise is an important parameter in developing the assembly plan. Therefore, the use of intelligent methods to plan the assembly process has been considered by many researchers. . The reinforcing learning approach has the potential to solve complex problems due to the use of experience gained from interacting with the environment and Has been successfully implemented in controlling many robotic tasks. However, due to the inherent complexity of the assembly, as well as...
Learning-based Control System Design for the Bipedal Running Robot and Development of a Two-layer Framework for Generating the Optimal Paths in Various Movement Maneuvers
, M.Sc. Thesis Sharif University of Technology ; Salarieh, Hassan (Supervisor)
Abstract
Foot movement is one of the most powerful and adaptable methods of movement in nature. Inspired by humans, the most intelligent creatures on earth, bipedal robots have many uses. In this research, a control method for running a bipedal robot has been designed. In the simulation part of the five-link model, the robot's motion equations for running and walking at different levels are extracted by the Lagrange method. In path generation, using the two-layer optimization method and holonomic and dynamic constraints, optimal paths are produced which are kinematically and dynamically possible (feasible). Additionally, path generation is facilitated by an invariant impact constraint to ensure the...
Robotic Arm Manipulation Learning from Demonstration based on Reinforcement Learning
, M.Sc. Thesis Sharif University of Technology ; Khodaygan, Saeed (Supervisor)
Abstract
The field of learning from demonstration is the field in which researchers seek to create methods by which a robot can learn and reproduce a skill simply by using the demonstration of the skill. One of the main drawbacks of learning from demonstration methods is their inability to improve the learned skills. To answer this question, the reinforcement learning method can be used. The reinforcement learning approach has the potential to improve the initial skill due to the use of the experience of interacting with the environment. In this project, the dynamic movement primitives algorithm is considered as the learning from demonstration method. The research approach is that first, the dynamic...
Brain Inspired Meta Reinforcement Learning Using Brain-Inspired Networks
, M.Sc. Thesis Sharif University of Technology ; Soleymani Baghshahi, Mahdih (Supervisor)
Abstract
Reinforcement learning is one of the most well-known learning paradigms in biological agents and one of the most used ones for solving plenty of problems. One of the reasons for this widespread use is the low demand for supervising signals. However, the sparsity of the reward signal causes increasing in sample complexity that needs for learning new tasks. This issue makes trouble in multi-task settings, specifically.One of the most promising approaches to learning new tasks by limited interaction with the environment is meta reinforcement learning. An approach in which fast adaption becomes possible by limiting hypothesis space and creating inductive biases by learning meta parameters....
Adaptive Maneuvers for Aircraft Conflict Resolution Using Learning Theory
, M.Sc. Thesis Sharif University of Technology ; Malaek, Mohammad Bagher (Supervisor)
Abstract
The problem of detection and resolution of aircraft collisions is very important due to the increasing demand for flights. Many algorithms have been developed in the past to increase automation in air traffic management and reduce the workload of air traffic controllers. These algorithms either have difficulty in generalizing to real problems or have high computational costs and do not correspond to the reality of the actual maneuvering characteristics of the aircraft performance. The aim of present study is to obtain dynamic maneuvers that are adaptive with reality and also optimal in terms of utilizing the capacity of flight sectors, so we propose Deep Reinforcement Learning(DRL) based on...
Optimal Control of a Quadcopter in Fast Descending Maneuvers Based on Reinforcement Learning
, M.Sc. Thesis Sharif University of Technology ; Fallah Rajabzadeh, Famida (Supervisor) ; Zohoor, Hassan (Supervisor) ; Nejat Pishkenari, Hossein (Co-Supervisor)
Abstract
Quadrotors have limitation in performing fast descent maneuvers due to Vortex Ring State (VRS) region which make quadrotor unstable. In order to avoid entering VRS, a velocity constraint considered which it should be satisfied during this maneuver to guarantee a safe and stable fast descending maneuver by quadrotor. The purpose of this thesis is to overcome limitation in speed space of quadrotor in order to reduce the time of fast descending maneuvers by using Reinforcement Learning Techniques. A new cascade controller proposed which using PID in inner loop as a low level controller and DDPG as one of reinforcement learning techniques in outer loop as high level controller in order to...
Design of a HEV’s Controller Using Learning-based Methods
, M.Sc. Thesis Sharif University of Technology ; Boroushaki, Mehrdad (Supervisor)
Abstract
Hybrid electric vehicles (HEV) are proving to be one of the most promising innovations in advanced transportation systems to reduce air pollution and fossil fuel consumption. EMS is one of the most vital aspects of the HEV powertrain system. This research aims to design an optimal EMS under the condition of meeting the goals of drivability control, fuel consumption reduction, and battery charge stability. The current EMS is based on the classical rule-based method derived from fuzzy logic, which guides to the suboptimal solution in episodic driving cycles. Previous experiences in implementing Reinforcement Learning (RL) suffer from late convergence, instability in tracking the driving...
Designing IoT-based Video/Audio Processing Systems
, M.Sc. Thesis Sharif University of Technology ; Gholampour, Iman (Supervisor) ; Haj Sadeghi, Khosrou (Supervisor)
Abstract
The use of IoT-based technologies is expanding in many areas today. The use of audio and video processing in IoT systems has been used as an alternative to human operators by increasing power and reducing processing costs. Due to the large volume of audio and video data and bandwidth limitations, complete data transfer to cloud processing servers is not cost-effective in terms of efficiency and energy consumption. As a result, the solution that has provided good results is to discharge these device tasks to the available clouds. In other words, the capacity of resources in the environment can be used to optimize the total latency of the system and energy consumption. In this dissertation, we...
A Novel Resource Allocation Algorithm in Edge Computing with Deep Reinforcement Learning
, M.Sc. Thesis Sharif University of Technology ; Movaghar, Ali (Supervisor)
Abstract
With the explosion of mobile smart devices, many computation intensive applications have emerged, such as interactive gaming and augmented reality. Mobile edge computing (EC) is put forward, as an extension of cloud computing, to meet the low-latency require- ments of the applications. In mobile edge computing systems, an edge node may have a high load when a large number of mobile devices offload their tasks to it. those offloaded tasks may experience large processing delay or even be dropped when their deadlines expire. Due to the uncertain load dynamics at the edge nodes, it is challenging for each device to determine its offloading decision (i.e., whether to offload or not, and which...
Deep Reinforcement Learning for Building Climate Control Using Weather Forecast Data
, M.Sc. Thesis Sharif University of Technology ; Rezaeizadeh, Amin (Supervisor)
Abstract
Buildings account for more than 30% of the world’s total energy consumption. Among building end-uses, air conditioning and in particular cooling systems have a major share of more than 50%. Therefore, design of optimal controllers for AC systems has become increasingly important. Classical and model-free control methods typically lack the ability to optimize energy consumption. On the other hand, model-based optimal control methods rely on precise modeling, which is difficult to acquire due to the complexity of the AC system dynamics.In recent years, deep reinforcement learning has become a popular choice for optimal control of systems with complex dynamics. In this thesis, a deep...
Safe Path Planning for Cooperative Mobile Robots Based on Deep Reinforcement Learning
, M.Sc. Thesis Sharif University of Technology ; Khodaygan, Saeed (Supervisor)
Abstract
Nowadays, with the remarkable development of the robotics industry, there is an increasing demand for mobile robots. Mobile robots can be deployed individually or in groups for various tasks such as autonomous warehouses, search and rescue operations, firefighting operations, and maintenance and repairs. It is evident that performing certain tasks, such as moving large and long objects or firefighting operations, is more efficient when robots are deployed cooperatively, and in some cases, these tasks cannot be accomplished by a single robot alone. Therefore, in recent years, the issue of path planning for cooperative robots has received significant attention. By cooperation, we mean that...
Multimodal Image Registration using Reinforcement Learning-based Methods
, M.Sc. Thesis Sharif University of Technology ; Fatemizadeh, Emadeddin (Supervisor)
Abstract
Image registration is the process of estimating and applying a spatial transformation to a moving image with the aim of spatially aligning it with a fixed image. This allows for the combination of images with complementary information, such as images with different modalities, acquisition times, and even coming from separate individuals, with the purpose of producing more information-rich results. Image registration is a crucial step in many medical applications, such as analyzing the growth and changes of tissue and tumors, preoperative planning, image-guided surgery, radiation therapy planning and various segmentation tasks. Reinforcement learning is a science and mathematical paradigm for...
A Reinforcement Learning Framework for Portfolio Management Problem Leveraging Stocks Historical Data And Their Correlation
, M.Sc. Thesis Sharif University of Technology ; Fazli, Mohammad Amin (Supervisor)
Abstract
Over the past few years, deep reinforcement learning(DRL) has been given a lot of attention in finance for portfolio management. With the help of experts’ signals and historical price data, we have developed a new reinforcement learning(RL) method. The use of experts’ signals in tandem with DRL has been used before in finance, but we believe this is the first time this method has been used to solve the financial portfolio management problem. As our agent, we used the Proximal Policy Optimization(PPO) algorithm to process the reward and take actions in the environment. Our framework comprises a convolutional network to aggregate signals, a convolutional network for historical price data, and...
Computation offloading strategy for autonomous vehicles
, Article 27th International Computer Conference, Computer Society of Iran, CSICC 2022, 23 February 2022 through 24 February 2022 ; 2022 ; 9781665480277 (ISBN) ; Karimian Aliabadi, S ; Entezari Maleki, R ; Sharif University of Technology
Institute of Electrical and Electronics Engineers Inc
2022
Abstract
Vehicular edge computing is a progressing technology which provides processing resources to the internet of vehicles using the edge servers deployed at roadside units. Vehicles take advantage by offloading their computationintensive tasks to this infrastructure. However, concerning time-sensitive applications and the high mobility of vehicles, cost-efficient task offloading is still a challenge. This paper establishes a computation offloading strategy based on deep Q-learning algorithm for vehicular edge computing networks. To jointly minimize the system cost including offloading failure rate and the total energy consumption of the offloading process, the vehicle tasks offloading problem is...
Firtual hardware-in-the-loop FMU CO-simulation based digital twins for heating, ventilation, and air-conditioning (HVAC) systems
, Article IEEE Transactions on Emerging Topics in Computational Intelligence ; 2022 , Pages 1-11 ; 2471285X (ISSN) ; Mohseni, S ; Zeitouni, M. J ; Parvaresh, A ; Fathollahi, A ; Gheisarnejad, M ; Khooban, M ; Sharif University of Technology
Institute of Electrical and Electronics Engineers Inc
2022
Abstract
In this paper, a novel self-adaptive control method based on a digital twin is developed and investigated for a multi-input multi-output (MIMO) nonlinear system, which is a heating, ventilation, and air-conditioning system. For this purpose, hardware-in-loop (HIL) and software-in-loop (SIL) are integrated to develop the digital twin control concept in a straightforward manner. A nonlinear integral backstepping (NIB) model-free control technique is integrated with the HIL (implemented as a physical controller) and SIL (implemented as a virtual controller) controllers to control the HVAC system without the need for dynamic feature identification. The main goal is to design the virtual...
An Application of Deep Reinforcement Learning in Novel Supply Chain Management Approaches for Inventory Control and Management of Perishable Supply Chain Network
, M.Sc. Thesis Sharif University of Technology ; Akhavan Niaki, Taghi (Supervisor)
Abstract
This study proposes a deep reinforcement learning approach to solve a perishable inventory allocation problem in a two-echelon supply chain. The inventory allocation problem is studied considering the stochastic nature of demand and supply. The examined supply chain includes two retailers and one distribution center (DC) under a vendor-managed inventory (VMI) system. This research aims to minimize the wastages and shortages occurring at the retailer's sites in the examined supply chain. With regard to continuous action space in the considered inventory allocation problem, the Advantage Actor-Critic algorithm is implemented to solve the problem. Numerical experiments are implemented on...