PyTorchKR Papers

최신 논문

총 1,872편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2012.03438

Selective Pseudo-Labeling with Reinforcement Learning for Semi-Supervised Domain Adaptation

Recent domain adaptation methods have demonstrated impressive improvement on unsupervised domain adaptation problems. However, in the semi-supervised domain adaptation (SSDA) setting where the target domain has a few labeled instances available, these methods can fail to improve performance. Inspired by the effectiveness of pseudo-labels in domain adaptation, we propose a reinforcement learning based selective pseudo-labeling method for semi-supervised domain adaptation. It is difficult for conventional pseudo-labeling methods to balance the correctness and representativeness of pseudo-labeled data. To address this limitation, we develop a deep Q-learning model to select both accurate and representative pseudo-labeled instances. Moreover, motivated by large margin loss's capacity on learning discriminative features with little data, we further propose a novel target margin loss for our base model training to improve its discriminability. Our proposed method is evaluated on several benchmark datasets for SSDA, and demonstrates superior performance to all the comparison methods.

arXiv:2012.01328

Hippocampal representations emerge when training recurrent neural networks on a memory dependent maze navigation task

Can neural networks learn goal-directed behaviour using similar strategies to the brain, by combining the relationships between the current state of the organism and the consequences of future actions? Recent work has shown that recurrent neural networks trained on goal based tasks can develop representations resembling those found in the brain, entorhinal cortex grid cells, for instance. Here we explore the evolution of the dynamics of their internal representations and compare this with experimental data. We observe that once a recurrent network is trained to learn the structure of its environment solely based on sensory prediction, an attractor based landscape forms in the network's representation, which parallels hippocampal place cells in structure and function. Next, we extend the predictive objective to include Q-learning for a reward task, where rewarding actions are dependent on delayed cue modulation. Mirroring experimental findings in hippocampus recordings in rodents performing the same task, this training paradigm causes nonlocal neural activity to sweep forward in space at decision points, anticipating the future path to a rewarded location. Moreover, prevalent choice and cue-selective neurons form in this network, again recapitulating experimental findings. Together, these results indicate that combining predictive, unsupervised learning of the structure of an environment with reinforcement learning can help understand the formation of hippocampus-like representations containing both spatial and task-relevant information.

arXiv:2008.05943

Multi-Agent Double Deep Q-Learning for Beamforming in mmWave MIMO Networks

Beamforming is one of the key techniques in millimeter wave (mmWave) multi-input multi-output (MIMO) communications. Designing appropriate beamforming not only improves the quality and strength of the received signal, but also can help reduce the interference, consequently enhancing the data rate. In this paper, we propose a distributed multi-agent double deep Q-learning algorithm for beamforming in mmWave MIMO networks, where multiple base stations (BSs) can automatically and dynamically adjust their beams to serve multiple highly-mobile user equipments (UEs). In the analysis, largest received power association criterion is considered for UEs, and a realistic channel model is taken into account. Simulation results demonstrate that the proposed learning-based algorithm can achieve comparable performance with respect to exhaustive search while operating at much lower complexity.

arXiv:2007.13965

Deep Reinforcement Learning for Dynamic Spectrum Sensing and Aggregation in Multi-Channel Wireless Networks

In this paper, the problem of dynamic spectrum sensing and aggregation is investigated in a wireless network containing N correlated channels, where these channels are occupied or vacant following an unknown joint 2-state Markov model. At each time slot, a single cognitive user with certain bandwidth requirement either stays idle or selects a segment comprising C (C < N) contiguous channels to sense. Then, the vacant channels in the selected segment will be aggregated for satisfying the user requirement. The user receives a binary feedback signal indicating whether the transmission is successful or not (i.e., ACK signal) after each transmission, and makes next decision based on the sensing channel states. Here, we aim to find a policy that can maximize the number of successful transmissions without interrupting the primary users (PUs). The problem can be considered as a partially observable Markov decision process (POMDP) due to without full observation of system environment. We implement a Deep Q-Network (DQN) to address the challenge of unknown system dynamics and computational expenses. The performance of DQN, Q-Learning, and the Improvident Policy with known system dynamics is evaluated through simulations. The simulation results show that DQN can achieve near-optimal performance among different system scenarios only based on partial observations and ACK signals.

arXiv:2006.08666

Runtime Adaptation in Wireless Sensor Nodes Using Structured Learning

Markov Decision Processes (MDPs) provide important capabilities for facilitating the dynamic adaptation and self-optimization of cyber physical systems at runtime. In recent years, this has primarily taken the form of Reinforcement Learning (RL) techniques that eliminate some MDP components for the purpose of reducing computational requirements. In this work, we show that recent advancements in Compact MDP Models (CMMs) provide sufficient cause to question this trend when designing wireless sensor network nodes. In this work, a novel CMM-based approach to designing self-aware wireless sensor nodes is presented and compared to Q-Learning, a popular RL technique. We show that a certain class of CPS nodes is not well served by RL methods, and contrast RL versus CMM methods in this context. Through both simulation and a prototype implementation, we demonstrate that CMM methods can provide significantly better runtime adaptation performance relative to Q-Learning, with comparable resource requirements.

arXiv:2006.06106

Privacy-Cost Management in Smart Meters with Mutual Information-Based Reinforcement Learning

The rapid development and expansion of the Internet of Things (IoT) paradigm has drastically increased the collection and exchange of data between sensors and systems, a phenomenon that raises serious privacy concerns in some domains. In particular, Smart Meters (SMs) share fine-grained electricity consumption of households with utility providers that can potentially violate users' privacy as sensitive information is leaked through the data. In order to enhance privacy, the electricity consumers can exploit the availability of physical resources such as a rechargeable battery (RB) to shape their power demand as dictated by a Privacy-Cost Management Unit (PCMU). In this paper, we present a novel method to learn the PCMU policy using Deep Reinforcement Learning (DRL). We adopt the mutual information (MI) between the user's demand load and the masked load seen by the power grid as a reliable and general privacy measure. Unlike previous studies, we model the whole temporal correlation in the data to learn the MI in its general form and use a neural network to estimate the MI-based reward signal to guide the PCMU learning process. This approach is combined with a model-free DRL algorithm known as the Deep Double Q-Learning (DDQL) method. The performance of the complete DDQL-MI algorithm is assessed empirically using an actual SMs dataset and compared with simpler privacy measures. Our results show significant improvements over state-of-the-art privacy-aware demand shaping methods.

arXiv:2006.05902

Q-greedyUCB: a New Exploration Policy for Adaptive and Resource-efficient Scheduling

This paper proposes a learning algorithm to find a scheduling policy that achieves an optimal delay-power trade-off in communication systems. Reinforcement learning (RL) is used to minimize the expected latency for a given energy constraint where the environments such as traffic arrival rates or channel conditions can change over time. For this purpose, this problem is formulated as an infinite-horizon Markov Decision Process (MDP) with constraints. To handle the constrained optimization problem, we adopt the Lagrangian relaxation technique to solve it. Then, we propose a variant of Q-learning, Q-greedyUCB that combines Q-learning for \emph{average} reward algorithm and Upper Confidence Bound (UCB) policy to solve this decision-making problem. We prove that the Q-greedyUCB algorithm is convergent through mathematical analysis. Simulation results show that Q-greedyUCB finds an optimal scheduling strategy, and is more efficient than Q-learning with the $\varepsilon$-greedy and Average-payoff RL algorithm in terms of the cumulative reward (i.e., the weighted sum of delay and energy) and the convergence speed. We also show that our algorithm can reduce the regret by up to 12% compared to the Q-learning with the $\varepsilon$-greedy and Average-payoff RL algorithm.

arXiv:2006.05574

Multi-Agent Reinforcement Learning in a Realistic Limit Order Book Market Simulation

Optimal order execution is widely studied by industry practitioners and academic researchers because it determines the profitability of investment decisions and high-level trading strategies, particularly those involving large volumes of orders. However, complex and unknown market dynamics pose significant challenges for the development and validation of optimal execution strategies. In this paper, we propose a model-free approach by training Reinforcement Learning (RL) agents in a realistic market simulation environment with multiple agents. First, we configure a multi-agent historical order book simulation environment for execution tasks built on an Agent-Based Interactive Discrete Event Simulation (ABIDES) [arXiv:1904.12066]. Second, we formulate the problem of optimal execution in an RL setting where an intelligent agent can make order execution and placement decisions based on market microstructure trading signals in High Frequency Trading (HFT). Third, we develop and train an RL execution agent using the Double Deep Q-Learning (DDQL) algorithm in the ABIDES environment. In some scenarios, our RL agent converges towards a Time-Weighted Average Price (TWAP) strategy. Finally, we evaluate the simulation with our RL agent by comparing it with a market replay simulation using real market Limit Order Book (LOB) data.

arXiv:2003.04946

Privacy-Cost Management in Smart Meters Using Deep Reinforcement Learning

Smart meters (SMs) play a pivotal rule in the smart grid by being able to report the electricity usage of consumers to the utility provider (UP) almost in real-time. However, this could leak sensitive information about the consumers to the UP or a third-party. Recent works have leveraged the availability of energy storage devices, e.g., a rechargeable battery (RB), in order to provide privacy to the consumers with minimal additional energy cost. In this paper, a privacy-cost management unit (PCMU) is proposed based on a model-free deep reinforcement learning algorithm, called deep double Q-learning (DDQL). Empirical results evaluated on actual SMs data are presented to compare DDQL with the state-of-the-art, i.e., classical Q-learning (CQL). Additionally, the performance of the method is investigated for two concrete cases where attackers aim to infer the actual demand load and the occupancy status of dwellings. Finally, an abstract information-theoretic characterization is provided.

arXiv:2002.08415

UAV Aided Search and Rescue Operation Using Reinforcement Learning

Owing to the enhanced flexibility in deployment and decreasing costs of manufacturing, the demand for unmanned aerial vehicles (UAVs) is expected to soar in the upcoming years. In this paper, we explore a UAV aided search and rescue~(SAR) operation in indoor environments, where the GPS signals might not be reliable. We consider a SAR scenario where the UAV tries to locate a victim trapped in an indoor environment by sensing the RF signals emitted from a smart device owned by the victim. To locate the victim as fast as possible, we leverage tools from reinforcement learning~(RL). Received signal strength~(RSS) at the UAV depends on the distance from the source, indoor shadowing, and fading parameters, and antenna radiation pattern of the receiver mounted on the UAV. To make our analysis more realistic, we model two indoor scenarios with different dimensions using commercial ray-tracing software. Then, the corresponding RSS values at each possible discrete UAV location are extracted and used in a Q-learning framework. Unlike the traditional location-based navigation approach that exploits GPS coordinates, our method uses the RSS to define the states and rewards of the RL algorithm. We compare the performance of the proposed method where directional and omnidirectional antennas are used. The results reveal that the use of directional antennas provides faster convergence rates than the omnidirectional antennas.

arXiv:1910.11673

Momentum-based Accelerated Q-learning

This paper studies accelerated algorithms for Q-learning. We propose an acceleration scheme by incorporating the historical iterates of the Q-function. The idea is conceptually inspired by the momentum-based acceleration methods in the optimization theory. Under finite state-action space settings, the proposed accelerated Q-learning algorithm provably converges to the global optimum with a rate of $\mathcal{O}(1/\sqrt{T})$. While sharing a comparable theoretic convergence rate with the existing Speedy Q-learning (SpeedyQ) algorithm, we numerically show that the proposed algorithm outperforms SpeedyQ via playing the FrozenLake grid world game. Furthermore, we generalize the acceleration scheme to the continuous state-action space case where function approximation of the Q-function is necessary. In this case, the algorithms are validated using commonly adopted testing problems in reinforcement learning, including two discrete-time linear quadratic regulation (LQR) problems from the Deepmind Control Suite, and the Atari 2600 games. Simulation results show that the proposed accelerated algorithms can improve the convergence performance compared with the vanilla Q-learning algorithm.

arXiv:1910.00107

Q-learning for POMDP: An application to learning locomotion gaits

This paper presents a Q-learning framework for learning optimal locomotion gaits in robotic systems modeled as coupled rigid bodies. Inspired by prevalence of periodic gaits in bio-locomotion, an open loop periodic input is assumed to (say) affect a nominal gait. The learning problem is to learn a new (modified) gait by using only partial noisy measurements of the state. The objective of learning is to maximize a given reward modeled as an objective function in optimal control settings. The proposed control architecture has three main components: (i) Phase modeling of dynamics by a single phase variable; (ii) A coupled oscillator feedback particle filter to represent the posterior distribution of the phase conditioned in the sensory measurements; and (iii) A Q-learning algorithm to learn the approximate optimal control law. The architecture is illustrated with the aid of a planar two-body system. The performance of the learning is demonstrated in a simulation environment.

arXiv:1906.08812

Cache-Aided NOMA Mobile Edge Computing: A Reinforcement Learning Approach

A novel non-orthogonal multiple access (NOMA) based cache-aided mobile edge computing (MEC) framework is proposed. For the purpose of efficiently allocating communication and computation resources to users' computation tasks requests, we propose a long-short-term memory (LSTM) network to predict the task popularity. Based on the predicted task popularity, a long-term reward maximization problem is formulated that involves a joint optimization of the task offloading decisions, computation resource allocation, and caching decisions. To tackle this challenging problem, a single-agent Q-learning (SAQ-learning) algorithm is invoked to learn a long-term resource allocation strategy. Furthermore, a Bayesian learning automata (BLA) based multi-agent Q-learning (MAQ-learning) algorithm is proposed for task offloading decisions. More specifically, a BLA based action select scheme is proposed for the agents in MAQ-learning to select the optimal action in every state. We prove that the BLA based action selection scheme is instantaneously self-correcting and the selected action is an optimal solution for each state. Extensive simulation results demonstrate that: 1) The prediction error of the proposed LSTMs based task popularity prediction decreases with increasing learning rate. 2) The proposed framework significantly outperforms the benchmarks like all local computing, all offloading computing, and non-cache computing. 3) The proposed BLA based MAQ-learning achieves an improved performance compared to conventional reinforcement learning algorithms.

arXiv:1903.06235

Learning Automata Based Q-learning for Content Placement in Cooperative Caching

An optimization problem of content placement in cooperative caching is formulated, with the aim of maximizing sum mean opinion score (MOS) of mobile users. Firstly, a supervised feed-forward back-propagation connectionist model based neural network (SFBC-NN) is invoked for user mobility and content popularity prediction. More particularly, practical data collected from GPS-tracker app on smartphones is tackled to test the accuracy of mobility prediction. Then, a learning automata-based Q-learning (LAQL) algorithm for cooperative caching is proposed, in which learning automata (LA) is invoked for Q-learning to obtain an optimal action selection in a random and stationary environment. It is proven that the LA-based action selection scheme is capable of enabling every state to select the optimal action with arbitrarily high probability if Q-learning is able to converge to the optimal Q value eventually. To characterize the performance of the proposed algorithms, the sum MOS of users is applied to define the reward function. Extensive simulations reveal that: 1) The prediction error of SFBC-NN lessen with the increase of iterations and nodes; 2) the proposed LAQL achieves significant performance improvement against traditional Q-learning; 3) the cooperative caching scheme is capable of outperforming non-cooperative caching and random caching of 3% and 4%.

arXiv:1810.10964

Finding the best design parameters for optical nanostructures using reinforcement learning

Recently, a novel machine learning model has emerged in the field of reinforcement learning known as deep Q-learning. This model is capable of finding the best possible solution in systems consisting of millions of choices, without ever experiencing it before, and has been used to beat the best human minds at complex games such as, Go and chess, which both have a huge number of possible decisions and outcomes for each move. With a human-level intelligence, it has been solved the problems that no other machine learning model could do before. Here, we show the steps needed for implementing this model on an optical problem. We investigated the colour generation by dielectric nanostructures and show that this model can find geometrical properties that can generate a much deeper red, green and blue colours compared to the ones found by human researchers. This technique can easily be extended to predict and find the best design parameters for other optical structures.

arXiv:1810.10408

Multi-Agent Reinforcement Learning Based Resource Allocation for UAV Networks

Unmanned aerial vehicles (UAVs) are capable of serving as aerial base stations (BSs) for providing both cost-effective and on-demand wireless communications. This article investigates dynamic resource allocation of multiple UAVs enabled communication networks with the goal of maximizing long-term rewards. More particularly, each UAV communicates with a ground user by automatically selecting its communicating users, power levels and subchannels without any information exchange among UAVs. To model the uncertainty of environments, we formulate the long-term resource allocation problem as a stochastic game for maximizing the expected rewards, where each UAV becomes a learning agent and each resource allocation solution corresponds to an action taken by the UAVs. Afterwards, we develop a multi-agent reinforcement learning (MARL) framework that each agent discovers its best strategy according to its local observations using learning. More specifically, we propose an agent-independent method, for which all agents conduct a decision algorithm independently but share a common structure based on Q-learning. Finally, simulation results reveal that: 1) appropriate parameters for exploitation and exploration are capable of enhancing the performance of the proposed MARL based resource allocation algorithm; 2) the proposed MARL algorithm provides acceptable performance compared to the case with complete information exchanges among UAVs. By doing so, it strikes a good tradeoff between performance gains and information exchange overheads.

arXiv:1806.06914

Distributional Advantage Actor-Critic

In traditional reinforcement learning, an agent maximizes the reward collected during its interaction with the environment by approximating the optimal policy through the estimation of value functions. Typically, given a state s and action a, the corresponding value is the expected discounted sum of rewards. The optimal action is then chosen to be the action a with the largest value estimated by value function. However, recent developments have shown both theoretical and experimental evidence of superior performance when value function is replaced with value distribution in context of deep Q learning [1]. In this paper, we develop a new algorithm that combines advantage actor-critic with value distribution estimated by quantile regression. We evaluated this new algorithm, termed Distributional Advantage Actor-Critic (DA2C or QR-A2C) on a variety of tasks, and observed it to achieve at least as good as baseline algorithms, and outperforming baseline in some tasks with smaller variance and increased stability.

arXiv:1805.07478

Algorithmic Trading with Fitted Q Iteration and Heston Model

We present the use of the fitted Q iteration in algorithmic trading. We show that the fitted Q iteration helps alleviate the dimension problem that the basic Q-learning algorithm faces in application to trading. Furthermore, we introduce a procedure including model fitting and data simulation to enrich training data as the lack of data is often a problem in realistic application. We experiment our method on both simulated environment that permits arbitrage opportunity and real-world environment by using prices of 450 stocks. In the former environment, the method performs well, implying that our method works in theory. To perform well in the real-world environment, the agents trained might require more training (iteration) and more meaningful variables with predictive value.

arXiv:2012.01100

Self-correcting Q-Learning

The Q-learning algorithm is known to be affected by the maximization bias, i.e. the systematic overestimation of action values, an important issue that has recently received renewed attention. Double Q-learning has been proposed as an efficient algorithm to mitigate this bias. However, this comes at the price of an underestimation of action values, in addition to increased memory requirements and a slower convergence. In this paper, we introduce a new way to address the maximization bias in the form of a "self-correcting algorithm" for approximating the maximum of an expected value. Our method balances the overestimation of the single estimator used in conventional Q-learning and the underestimation of the double estimator used in Double Q-learning. Applying this strategy to Q-learning results in Self-correcting Q-learning. We show theoretically that this new algorithm enjoys the same convergence guarantees as Q-learning while being more accurate. Empirically, it performs better than Double Q-learning in domains with rewards of high variance, and it even attains faster convergence than Q-learning in domains with rewards of zero or low variance. These advantages transfer to a Deep Q Network implementation that we call Self-correcting DQN and which outperforms regular DQN and Double DQN on several tasks in the Atari 2600 domain.

arXiv:2012.00065

Deep reinforcement learning with a particle dynamics environment applied to emergency evacuation of a room with obstacles

A very successful model for simulating emergency evacuation is the social-force model. At the heart of the model is the self-driven force that is applied to an agent and is directed towards the exit. However, it is not clear if the application of this force results in optimal evacuation, especially in complex environments with obstacles. Here, we develop a deep reinforcement learning algorithm in association with the social force model to train agents to find the fastest evacuation path. During training, we penalize every step of an agent in the room and give zero reward at the exit. We adopt the Dyna-Q learning approach. We first show that in the case of a room without obstacles the resulting self-driven force points directly towards the exit as in the social force model and that the median exit time intervals calculated using the two methods are not significantly different. Then, we investigate evacuation of a room with one obstacle and one exit. We show that our method produces similar results with the social force model when the obstacle is convex. However, in the case of concave obstacles, which sometimes can act as traps for agents governed purely by the social force model and prohibit complete room evacuation, our approach is clearly advantageous since it derives a policy that results in object avoidance and complete room evacuation without additional assumptions. We also study evacuation of a room with multiple exits. We show that agents are able to evacuate efficiently from the nearest exit through a shared network trained for a single agent. Finally, we test the robustness of the Dyna-Q learning approach in a complex environment with multiple exits and obstacles. Overall, we show that our model can efficiently simulate emergency evacuation in complex environments with multiple room exits and obstacles where it is difficult to obtain an intuitive rule for fast evacuation.