Equipped with the trained environmental dynamics, model-based offline reinforcement learning (RL) algorithms can often successfully learn good policies from fixed-sized datasets, even some datasets with poor quality. Unfortunately, however, it can not be guaranteed that the generated samples from the trained dynamics model are reliable (e.g., some synthetic samples may lie outside of the support region of the static dataset). To address this issue, we propose Trajectory Truncation with Uncertainty (TATU), which adaptively truncates the synthetic trajectory if the accumulated uncertainty along the trajectory is too large. We theoretically show the performance bound of TATU to justify its benefits. To empirically show the advantages of TATU, we first combine it with two classical model-based offline RL algorithms, MOPO and COMBO. Furthermore, we integrate TATU with several off-the-shelf model-free offline RL algorithms, e.g., BCQ. Experimental results on the D4RL benchmark show that TATU significantly improves their performance, often by a large margin. Code is available here.

Needle picking is a challenging manipulation task in robot-assisted surgery due to the characteristics of small slender shapes of needles, needles' variations in shapes and sizes, and demands for millimeter-level control. Prior works, heavily relying on the prior of needles (e.g., geometric models), are hard to scale to unseen needles' variations. In this paper, we present the first end-to-end learning method to train deep visuomotor policy for needle picking. Concretely, we propose DreamerfD to maximally leverage demonstrations to improve the learning efficiency of a state-of-the-art model-based reinforcement learning method, DreamerV2; Since Variational Auto-Encoder (VAE) in DreamerV2 is difficult to scale to high-resolution images, we propose Dynamic Spotlight Adaptation to represent control-related visual signals in a low-resolution image space; Virtual Clutch is also proposed to reduce performance degradation due to significant error between prior and posterior encoded states at the beginning of a rollout. We conducted extensive experiments in simulation to evaluate the performance, robustness, in-domain variation adaptation, and effectiveness of individual components of our method. Our method, trained by 8k demonstration timesteps and 140k online policy timesteps, can achieve a remarkable success rate of 80%. Furthermore, our method effectively demonstrated its superiority in generalization to unseen in-domain variations including needle variations and image disturbance, highlighting its robustness and versatility. Codes and videos are available at https://sites.google.com/view/DreamerfD.

More and more stock trading strategies are constructed using deep reinforcement learning (DRL) algorithms, but DRL methods originally widely used in the gaming community are not directly adaptable to financial data with low signal-to-noise ratios and unevenness, and thus suffer from performance shortcomings. In this paper, to capture the hidden information, we propose a DRL based stock trading system using cascaded LSTM, which first uses LSTM to extract the time-series features from stock daily data, and then the features extracted are fed to the agent for training, while the strategy functions in reinforcement learning also use another LSTM for training. Experiments in DJI in the US market and SSE50 in the Chinese stock market show that our model outperforms previous baseline models in terms of cumulative returns and Sharp ratio, and this advantage is more significant in the Chinese stock market, a merging market. It indicates that our proposed method is a promising way to build a automated stock trading system.

The paper applies reinforcement learning to novel Internet of Thing configurations. Our analysis of inaudible attacks on voice-activated devices confirms the alarming risk factor of 7.6 out of 10, underlining significant security vulnerabilities scored independently by NIST National Vulnerability Database (NVD). Our baseline network model showcases a scenario in which an attacker uses inaudible voice commands to gain unauthorized access to confidential information on a secured laptop. We simulated many attack scenarios on this baseline network model, revealing the potential for mass exploitation of interconnected devices to discover and own privileged information through physical access without adding new hardware or amplifying device skills. Using Microsoft's CyberBattleSim framework, we evaluated six reinforcement learning algorithms and found that Deep-Q learning with exploitation proved optimal, leading to rapid ownership of all nodes in fewer steps. Our findings underscore the critical need for understanding non-conventional networks and new cybersecurity measures in an ever-expanding digital landscape, particularly those characterized by mobile devices, voice activation, and non-linear microphones susceptible to malicious actors operating stealth attacks in the near-ultrasound or inaudible ranges. By 2024, this new attack surface might encompass more digital voice assistants than people on the planet yet offer fewer remedies than conventional patching or firmware fixes since the inaudible attacks arise inherently from the microphone design and digital signal processing. Voice-activated devices, such as digital voice assistants, have experienced rapid proliferation in recent years.

Dynamic adaptive streaming over HTTP (DASH) has been widely used in video streaming recently. In DASH, the client downloads video chunks in order from a server. The rate adaptation function at the video client enhances the user's quality-of-experience (QoE) by choosing a suitable quality level for each video chunk to download based on the network condition. Today networks such as content delivery networks, edge caching networks, content-centric networks,... usually replicate video contents on multiple cache nodes. We study video streaming from multiple sources in this work. In multi-source streaming, video chunks may arrive out of order due to different conditions of the network paths. Hence, to guarantee a high QoE, the video client needs not only rate adaptation but also chunk scheduling. Reinforcement learning (RL) has emerged as the state-of-the-art control method in various fields in recent years. This paper proposes two algorithms for streaming from multiple sources: RL-based adaptation with greedy scheduling (RLAGS) and RL-based adaptation and scheduling (RLAS). We also build a simulation environment for training and evaluating. The efficiency of the proposed algorithms is proved via extensive simulations with real-trace data.

The core challenge of offline reinforcement learning (RL) is dealing with the (potentially catastrophic) extrapolation error induced by the distribution shift between the history dataset and the desired policy. A large portion of prior work tackles this challenge by implicitly/explicitly regularizing the learning policy towards the behavior policy, which is hard to estimate reliably in practice. In this work, we propose to regularize towards the Q-function of the behavior policy instead of the behavior policy itself, under the premise that the Q-function can be estimated more reliably and easily by a SARSA-style estimate and handles the extrapolation error more straightforwardly. We propose two algorithms taking advantage of the estimated Q-function through regularizations, and demonstrate they exhibit strong performance on the D4RL benchmarks.

Traditional federated learning uses the number of samples to calculate the weights of each client model and uses this fixed weight value to fusion the global model. However, in practical scenarios, each client's device and data heterogeneity leads to differences in the quality of each client's model. Thus the contribution to the global model is not wholly determined by the sample size. In addition, if clients intentionally upload low-quality or malicious models, using these models for aggregation will lead to a severe decrease in global model accuracy. Traditional federated learning algorithms do not address these issues. To solve this probelm, we propose FedDRL, a model fusion approach using reinforcement learning based on a two staged approach. In the first stage, Our method could filter out malicious models and selects trusted client models to participate in the model fusion. In the second stage, the FedDRL algorithm adaptively adjusts the weights of the trusted client models and aggregates the optimal global model. We also define five model fusion scenarios and compare our method with two baseline algorithms in those scenarios. The experimental results show that our algorithm has higher reliability than other algorithms while maintaining accuracy.

Any autonomous controller will be unsafe in some situations. The ability to quantitatively identify when these unsafe situations are about to occur is crucial for drawing timely human oversight in, e.g., freight transportation applications. In this work, we demonstrate that the true criticality of an agent's situation can be robustly defined as the mean reduction in reward given some number of random actions. Proxy criticality metrics that are computable in real-time (i.e., without actually simulating the effects of random actions) can be compared to the true criticality, and we show how to leverage these proxy metrics to generate safety margins, which directly tie the consequences of potentially incorrect actions to an anticipated loss in overall performance. We evaluate our approach on learned policies from APE-X and A3C within an Atari environment, and demonstrate how safety margins decrease as agents approach failure states. The integration of safety margins into programs for monitoring deployed agents allows for the real-time identification of potentially catastrophic situations.

The rapid expansion of global cloud wide-area networks (WANs) has posed a challenge for commercial optimization engines to efficiently solve network traffic engineering (TE) problems at scale. Existing acceleration strategies decompose TE optimization into concurrent subproblems but realize limited parallelism due to an inherent tradeoff between run time and allocation performance. We present Teal, a learning-based TE algorithm that leverages the parallel processing power of GPUs to accelerate TE control. First, Teal designs a flow-centric graph neural network (GNN) to capture WAN connectivity and network flows, learning flow features as inputs to downstream allocation. Second, to reduce the problem scale and make learning tractable, Teal employs a multi-agent reinforcement learning (RL) algorithm to independently allocate each traffic demand while optimizing a central TE objective. Finally, Teal fine-tunes allocations with ADMM (Alternating Direction Method of Multipliers), a highly parallelizable optimization algorithm for reducing constraint violations such as overutilized links. We evaluate Teal using traffic matrices from Microsoft's WAN. On a large WAN topology with >1,700 nodes, Teal generates near-optimal flow allocations while running several orders of magnitude faster than the production optimization engine. Compared with other TE acceleration schemes, Teal satisfies 6--32% more traffic demand and yields 197--625x speedups.

A central issue lying at the heart of online reinforcement learning (RL) is data efficiency. While a number of recent works achieved asymptotically minimal regret in online RL, the optimality of these results is only guaranteed in a ``large-sample'' regime, imposing enormous burn-in cost in order for their algorithms to operate optimally. How to achieve minimax-optimal regret without incurring any burn-in cost has been an open problem in RL theory. We settle this problem for the context of finite-horizon inhomogeneous Markov decision processes. Specifically, we prove that a modified version of Monotonic Value Propagation (MVP), a model-based algorithm proposed by \cite{zhang2020reinforcement}, achieves a regret on the order of (modulo log factors) \begin{equation*} \min\big\{ \sqrt{SAH^3K}, \,HK \big\}, \end{equation*} where $S$ is the number of states, $A$ is the number of actions, $H$ is the planning horizon, and $K$ is the total number of episodes. This regret matches the minimax lower bound for the entire range of sample size $K\geq 1$, essentially eliminating any burn-in requirement. It also translates to a PAC sample complexity (i.e., the number of episodes needed to yield $\varepsilon$-accuracy) of $\frac{SAH^3}{\varepsilon^2}$ up to log factor, which is minimax-optimal for the full $\varepsilon$-range. Further, we extend our theory to unveil the influences of problem-dependent quantities like the optimal value/cost and certain variances. The key technical innovation lies in the development of a new regret decomposition strategy and a novel analysis paradigm to decouple complicated statistical dependency -- a long-standing challenge facing the analysis of online RL in the sample-hungry regime.