A primary requirement for any reinforcement learning method is that it should produce policies that improve upon the initial guess. In this work, we show that the widely used Deep Q-Network (DQN) fails to satisfy this minimal criterion -- even when it gets to see all possible states and actions infinitely often (a condition under which tabular Q-learning is guaranteed to converge to the optimal Q-value function). Our specific contributions are twofold. First, we numerically show that DQN often returns a policy that performs worse than the initial one. Second, we offer a theoretical explanation for this phenomenon in linear DQN, a simplified version of DQN that uses linear function approximation in place of neural networks while retaining the other key components such as $\epsilon$-greedy exploration, experience replay, and target network. Using tools from differential inclusion theory, we prove that the limit points of linear DQN correspond to fixed points of projected Bellman operators. Crucially, we show that these fixed points need not relate to optimal -- or even near-optimal -- policies, thus explaining linear DQN's sub-optimal behaviors. We also give a scenario where linear DQN always identifies the worst policy. Our work fills a longstanding gap in understanding the convergence behaviors of Q-learning with function approximation and $\epsilon$-greedy exploration.
Multi-armed bandit (MAB) is a classic model for understanding the exploration-exploitation trade-off. The traditional MAB model for recommendation systems assumes the user stays in the system for the entire learning horizon. In new online education platforms such as ALEKS or new video recommendation systems such as TikTok and YouTube Shorts, the amount of time a user spends on the app depends on how engaging the recommended contents are. Users may temporarily leave the system if the recommended items cannot engage the users. To understand the exploration, exploitation, and engagement in these systems, we propose a new model, called MAB-A where "A" stands for abandonment and the abandonment probability depends on the current recommended item and the user's past experience (called state). We propose two algorithms, ULCB and KL-ULCB, both of which do more exploration (being optimistic) when the user likes the previous recommended item and less exploration (being pessimistic) when the user does not like the previous item. We prove that both ULCB and KL-ULCB achieve logarithmic regret, $O(\log K)$, where $K$ is the number of visits (or episodes). Furthermore, the regret bound under KL-ULCB is asymptotically sharp. We also extend the proposed algorithms to the general-state setting. Simulation results confirm our theoretical analysis and show that the proposed algorithms have significantly lower regrets than the traditional UCB and KL-UCB, and Q-learning-based algorithms.
This paper presents a novel deep reinforcement learning-based resource allocation technique for the multi-agent environment presented by a cognitive radio network where the interactions of the agents during learning may lead to a non-stationary environment. The resource allocation technique presented in this work is distributed, not requiring coordination with other agents. It is shown by considering aspects specific to deep reinforcement learning that the presented algorithm converges in an arbitrarily long time to equilibrium policies in a non-stationary multi-agent environment that results from the uncoordinated dynamic interaction between radios through the shared wireless environment. Simulation results show that the presented technique achieves a faster learning performance compared to an equivalent table-based Q-learning algorithm and is able to find the optimal policy in 99% of cases for a sufficiently long learning time. In addition, simulations show that our DQL approach requires less than half the number of learning steps to achieve the same performance as an equivalent table-based implementation. Moreover, it is shown that the use of a standard single-agent deep reinforcement learning approach may not achieve convergence when used in an uncoordinated interacting multi-radio scenario
In the survey we consider the case studies on sales time series forecasting, the deep learning approach for forecasting non-stationary time series using time trend correction, dynamic price and supply optimization using Q-learning, Bitcoin price modeling, COVID-19 spread impact on stock market, using social networks signals in analytics. The use of machine learning and Bayesian inference in predictive analytics has been analyzed.
Gradient descent or its variants are popular in training neural networks. However, in deep Q-learning with neural network approximation, a type of reinforcement learning, gradient descent (also known as Residual Gradient (RG)) is barely used to solve Bellman residual minimization problem. On the contrary, Temporal Difference (TD), an incomplete gradient descent method prevails. In this work, we perform extensive experiments to show that TD outperforms RG, that is, when the training leads to a small Bellman residual error, the solution found by TD has a better policy and is more robust against the perturbation of neural network parameters. We further use experiments to reveal a key difference between reinforcement learning and supervised learning, that is, a small Bellman residual error can correspond to a bad policy in reinforcement learning while the test loss function in supervised learning is a standard index to indicate the performance. We also empirically examine that the missing term in TD is a key reason why RG performs badly. Our work shows that the performance of a deep Q-learning solution is closely related to the training dynamics and how an incomplete gradient descent method can find a good policy is interesting for future study.
With the rapid growth of memory and computing power, datasets are becoming increasingly complex and imbalanced. This is especially severe in the context of clinical data, where there may be one rare event for many cases in the majority class. We introduce an imbalanced classification framework, based on reinforcement learning, for training extremely imbalanced data sets, and extend it for use in multi-class settings. We combine dueling and double deep Q-learning architectures, and formulate a custom reward function and episode-training procedure, specifically with the added capability of handling multi-class imbalanced training. Using real-world clinical case studies, we demonstrate that our proposed framework outperforms current state-of-the-art imbalanced learning methods, achieving more fair and balanced classification, while also significantly improving the prediction of minority classes.
Momentum and mean reversion trading strategies have opposite characteristics. The former is generally better with trending assets, and the latter is generally better with mean reverting assets. Using the Hurst exponent, which classifies time series as trending or mean reverting, we attempt to trade with each strategy when it is advantageous to generate higher returns on average. We ultimately find that trading with the Hurst exponent can achieve higher returns, but it also comes at a higher risk. Finally, we consider limitations of our study and propose a method using Q-learning to improve our strategy and implementation of individual algorithms.
Accelerating artificial intelligence by photonics is an active field of study aiming to exploit the unique properties of photons. Reinforcement learning is an important branch of machine learning, and photonic decision-making principles have been demonstrated with respect to the multi-armed bandit problems. However, reinforcement learning could involve a massive number of states, unlike previously demonstrated bandit problems where the number of states is only one. Q-learning is a well-known approach in reinforcement learning that can deal with many states. The architecture of Q-learning, however, does not fit well photonic implementations due to its separation of update rule and the action selection. In this study, we organize a new architecture for multi-state reinforcement learning as a parallel array of bandit problems in order to benefit from photonic decision-makers, which we call parallel bandit architecture for reinforcement learning or PBRL in short. Taking a cart-pole balancing problem as an instance, we demonstrate that PBRL adapts to the environment in fewer time steps than Q-learning. Furthermore, PBRL yields faster adaptation when operated with a chaotic laser time series than the case with uniformly distributed pseudorandom numbers where the autocorrelation inherent in the laser chaos provides a positive effect. We also find that the variety of states that the system undergoes during the learning phase exhibits completely different properties between PBRL and Q-learning. The insights obtained through the present study are also beneficial for existing computing platforms, not just photonic realizations, in accelerating performances by the PBRL algorithms and correlated random sequences.
Reinforcement Learning (RL) has opened up new opportunities to enhance existing smart systems that generally include a complex decision-making process. However, modern RL algorithms, e.g., Deep Q-Networks (DQN), are based on deep neural networks, resulting in high computational costs. In this paper, we propose QHD, an off-policy value-based Hyperdimensional Reinforcement Learning, that mimics brain properties toward robust and real-time learning. QHD relies on a lightweight brain-inspired model to learn an optimal policy in an unknown environment. On both desktop and power-limited embedded platforms, QHD achieves significantly better overall efficiency than DQN while providing higher or comparable rewards. QHD is also suitable for highly-efficient reinforcement learning with great potential for online and real-time learning. Our solution supports a small experience replay batch size that provides 12.3 times speedup compared to DQN while ensuring minimal quality loss. Our evaluation shows QHD capability for real-time learning, providing 34.6 times speedup and significantly better quality of learning than DQN.
Humans are capable of adjusting to changing environments flexibly and quickly. Empirical evidence has revealed that representation learning plays a crucial role in endowing humans with such a capability. Inspired by this observation, we study representation learning in the sequential decision-making scenario with contextual changes. We propose an online algorithm that is able to learn and transfer context-dependent representations and show that it significantly outperforms the existing ones that do not learn representations adaptively. As a case study, we apply our algorithm to the Wisconsin Card Sorting Task, a well-established test for the mental flexibility of humans in sequential decision-making. By comparing our algorithm with the standard Q-learning and Deep-Q learning algorithms, we demonstrate the benefits of adaptive representation learning.
Q-learning is known as one of the fundamental reinforcement learning (RL) algorithms. Its convergence has been the focus of extensive research over the past several decades. Recently, a new finitetime error bound and analysis for Q-learning was introduced using a switching system framework. This approach views the dynamics of Q-learning as a discrete-time stochastic switching system. The prior study established a finite-time error bound on the averaged iterates using Lyapunov functions, offering further insights into Q-learning. While valuable, the analysis focuses on error bounds of the averaged iterate, which comes with the inherent disadvantages: it necessitates extra averaging steps, which can decelerate the convergence rate. Moreover, the final iterate, being the original format of Q-learning, is more commonly used and is often regarded as a more intuitive and natural form in the majority of iterative algorithms. In this paper, we present a finite-time error bound on the final iterate of Q-learning based on the switching system framework. The proposed error bounds have different features compared to the previous works, and cover different scenarios. Finally, we expect that the proposed results provide additional insights on Q-learning via connections with discrete-time switching systems, and can potentially present a new template for finite-time analysis of more general RL algorithms.
We study the action generalization ability of deep Q-learning in discrete action spaces. Generalization is crucial for efficient reinforcement learning (RL) because it allows agents to use knowledge learned from past experiences on new tasks. But while function approximation provides deep RL agents with a natural way to generalize over state inputs, the same generalization mechanism does not apply to discrete action outputs. And yet, surprisingly, our experiments indicate that Deep Q-Networks (DQN), which use exactly this type of function approximator, are still able to achieve modest action generalization. Our main contribution is twofold: first, we propose a method of evaluating action generalization using expert knowledge of action similarity, and empirically confirm that action generalization leads to faster learning; second, we characterize the action-generalization gap (the difference in learning performance between DQN and the expert) in different domains. We find that DQN can indeed generalize over actions in several simple domains, but that its ability to do so decreases as the action space grows larger.
Building on our recent work on {\em neuromimetic control theory}, new results on resilience and neuro-inspired quantization are reported. The term neuromimetic refers to the models having features that are characteristic of the neurobiology of biological motor control. As in previous work, the focus is on what we call {\em overcomplete} linear systems that are characterized by larger numbers of input and output channels than the dimensions of the state. The specific contributions of the present paper include a proposed {\em resilient} observer whose operation tolerates output channel intermittency and even complete dropouts. Tying these ideas together with our previous work on resilient stability, a resilient separation principle is established. We also propose a {\em principled quantization} in which control signals are encoded as simple discrete inputs which act collectively through the many channels of input that are the hallmarks of the overcomplete models. Aligned with the neuromimetic paradigm, an {\em emulation} problem is proposed and this in turn defines an optimal quantization problem. Several possible solutions are discussed including direct combinatorial optimization, a Hebbian-like iterative learning algorithm, and a deep Q-learning (DQN) approach. For the problems being considered, machine learning approaches to optimization provide valuable insights regarding comparisons between optimal and nearby suboptimal solutions. These are useful in understanding the kinds of resilience to intermittency and channel dropouts that were earlier demonstrated for continuous systems.
Actor-critic Reinforcement Learning (RL) algorithms have achieved impressive performance in continuous control tasks. However, they still suffer two nontrivial obstacles, i.e., low sample efficiency and overestimation bias. To this end, we propose Simultaneous Double Q-learning with Conservative Advantage Learning (SDQ-CAL). Our SDQ-CAL boosts the Double Q-learning for off-policy actor-critic RL based on a modification of the Bellman optimality operator with Advantage Learning. Specifically, SDQ-CAL improves sample efficiency by modifying the reward to facilitate the distinction from experience between the optimal actions and the others. Besides, it mitigates the overestimation issue by updating a pair of critics simultaneously upon double estimators. Extensive experiments reveal that our algorithm realizes less biased value estimation and achieves state-of-the-art performance in a range of continuous control benchmark tasks. We release the source code of our method at: \url{https://github.com/LQNew/SDQ-CAL}.
We investigate the feasibility of deploying Deep-Q based deep reinforcement learning agents to job-shop scheduling problems in the context of modular production facilities, using discrete event simulations for the environment. These environments are comprised of a source and sink for the parts to be processed, as well as (several) workstations. The agents are trained to schedule automated guided vehicles to transport the parts back and forth between those stations in an optimal fashion. Starting from a very simplistic setup, we increase the complexity of the environment and compare the agents' performances with well established heuristic approaches, such as first-in-first-out based agents, cost tables and a nearest-neighbor approach. We furthermore seek particular configurations of the environments in which the heuristic approaches struggle, to investigate to what degree the Deep-Q agents are affected by these challenges. We find that Deep-Q based agents show comparable performance as the heuristic baselines. Furthermore, our findings suggest that the DRL agents exhibit an increased robustness to noise, as compared to the conventional approaches. Overall, we find that DRL agents constitute a valuable approach for this type of scheduling problems.
The coupled problem of hydrodynamics and solute transport for the Najafi-Golestanian three-sphere swimmer is studied, with the Reynolds number set to zero and P\'eclet numbers (Pe) ranging from 0.06 to 60. The adopted method is the numerical simulation of the problem with a finite element code based upon the FEniCS library. For the swimmer executing the optimal locomotion gait, we report the Sherwood number as a function of Pe in homogeneous fluids and confirm that little gain in solute flux is achieved by swimming unless Pe is significantly larger than 10. We also consider the swimmer as an learning agent moving inside a fluid that has a concentration gradient. The outcomes of Q-learning processes show that learning locomotion (with the displacement as reward) is significantly easier than learning chemotaxis (with the increase of solute flux as reward). The chemotaxis problem, even at low Pe, has a varying environment that renders learning more difficult. Further, the learning difficulty increases severely with the P\'eclet number. The results demonstrate the challenges that natural and artificial swimmers need to overcome to migrate efficiently when exposed to chemical inhomogeneities.
In reinforcement learning (RL), it is challenging to learn directly from high-dimensional observations, where data augmentation has recently been shown to remedy this via encoding invariances from raw pixels. Nevertheless, we empirically find that not all samples are equally important and hence simply injecting more augmented inputs may instead cause instability in Q-learning. In this paper, we approach this problem systematically by developing a model-agnostic Contrastive-Curiosity-Driven Learning Framework (CCLF), which can fully exploit sample importance and improve learning efficiency in a self-supervised manner. Facilitated by the proposed contrastive curiosity, CCLF is capable of prioritizing the experience replay, selecting the most informative augmented inputs, and more importantly regularizing the Q-function as well as the encoder to concentrate more on under-learned data. Moreover, it encourages the agent to explore with a curiosity-based reward. As a result, the agent can focus on more informative samples and learn representation invariances more efficiently, with significantly reduced augmented inputs. We apply CCLF to several base RL algorithms and evaluate on the DeepMind Control Suite, Atari, and MiniGrid benchmarks, where our approach demonstrates superior sample efficiency and learning performances compared with other state-of-the-art methods.
This paper tackles the problem of learning value functions from undirected state-only experience (state transitions without action labels i.e. (s,s',r) tuples). We first theoretically characterize the applicability of Q-learning in this setting. We show that tabular Q-learning in discrete Markov decision processes (MDPs) learns the same value function under any arbitrary refinement of the action space. This theoretical result motivates the design of Latent Action Q-learning or LAQ, an offline RL method that can learn effective value functions from state-only experience. Latent Action Q-learning (LAQ) learns value functions using Q-learning on discrete latent actions obtained through a latent-variable future prediction model. We show that LAQ can recover value functions that have high correlation with value functions learned using ground truth actions. Value functions learned using LAQ lead to sample efficient acquisition of goal-directed behavior, can be used with domain-specific low-level controllers, and facilitate transfer across embodiments. Our experiments in 5 environments ranging from 2D grid world to 3D visual navigation in realistic environments demonstrate the benefits of LAQ over simpler alternatives, imitation learning oracles, and competing methods.
We study the problem of reinforcement learning for a task encoded by a reward machine. The task is defined over a set of properties in the environment, called atomic propositions, and represented by Boolean variables. One unrealistic assumption commonly used in the literature is that the truth values of these propositions are accurately known. In real situations, however, these truth values are uncertain since they come from sensors that suffer from imperfections. At the same time, reward machines can be difficult to model explicitly, especially when they encode complicated tasks. We develop a reinforcement-learning algorithm that infers a reward machine that encodes the underlying task while learning how to execute it, despite the uncertainties of the propositions' truth values. In order to address such uncertainties, the algorithm maintains a probabilistic estimate about the truth value of the atomic propositions; it updates this estimate according to new sensory measurements that arrive from the exploration of the environment. Additionally, the algorithm maintains a hypothesis reward machine, which acts as an estimate of the reward machine that encodes the task to be learned. As the agent explores the environment, the algorithm updates the hypothesis reward machine according to the obtained rewards and the estimate of the atomic propositions' truth value. Finally, the algorithm uses a Q-learning procedure for the states of the hypothesis reward machine to determine the policy that accomplishes the task. We prove that the algorithm successfully infers the reward machine and asymptotically learns a policy that accomplishes the respective task.
Deep neural networks (DNN) can approximate value functions or policies for reinforcement learning, which makes the reinforcement learning algorithms more powerful. However, some DNNs, such as convolutional neural networks (CNN), cannot extract enough information or take too long to obtain enough features from the inputs under specific circumstances of reinforcement learning. For example, the input data of Google Research Football, a reinforcement learning environment which trains agents to play football, is the small map of players' locations. The information is contained not only in the coordinates of players, but also in the relationships between different players. CNNs can neither extract enough information nor take too long to train. To address this issue, this paper proposes a deep q-learning network (DQN) with a graph neural network (GNN) as its model. The GNN transforms the input data into a graph which better represents the football players' locations so that it extracts more information of the interactions between different players. With two GNNs to approximate its local and target value functions, this DQN allows players to learn from their experience by using value functions to see the prospective value of each intended action. The proposed model demonstrated the power of GNN in the football game by outperforming other DRL models with significantly fewer steps.