Hierarchical Reinforcement Learning (HRL) algorithms have been demonstrated to perform well on high-dimensional decision making and robotic control tasks. However, because they solely optimize for rewards, the agent tends to search the same space redundantly. This problem reduces the speed of learning and achieved reward. In this work, we present an Off-Policy HRL algorithm that maximizes entropy for efficient exploration. The algorithm learns a temporally abstracted low-level policy and is able to explore broadly through the addition of entropy to the high-level. The novelty of this work is the theoretical motivation of adding entropy to the RL objective in the HRL setting. We empirically show that the entropy can be added to both levels if the Kullback-Leibler (KL) divergence between consecutive updates of the low-level policy is sufficiently small. We performed an ablative study to analyze the effects of entropy on hierarchy, in which adding entropy to high-level emerged as the most desirable configuration. Furthermore, a higher temperature in the low-level leads to Q-value overestimation and increases the stochasticity of the environment that the high-level operates on, making learning more challenging. Our method, SHIRO, surpasses state-of-the-art performance on a range of simulated robotic control benchmark tasks and requires minimal tuning.

In recent years, reinforcement learning (RL) has become increasingly successful in its application to science and the process of scientific discovery in general. However, while RL algorithms learn to solve increasingly complex problems, interpreting the solutions they provide becomes ever more challenging. In this work, we gain insights into an RL agent's learned behavior through a post-hoc analysis based on sequence mining and clustering. Specifically, frequent and compact subroutines, used by the agent to solve a given task, are distilled as gadgets and then grouped by various metrics. This process of gadget discovery develops in three stages: First, we use an RL agent to generate data, then, we employ a mining algorithm to extract gadgets and finally, the obtained gadgets are grouped by a density-based clustering algorithm. We demonstrate our method by applying it to two quantum-inspired RL environments. First, we consider simulated quantum optics experiments for the design of high-dimensional multipartite entangled states where the algorithm finds gadgets that correspond to modern interferometer setups. Second, we consider a circuit-based quantum computing environment where the algorithm discovers various gadgets for quantum information processing, such as quantum teleportation. This approach for analyzing the policy of a learned agent is agent and environment agnostic and can yield interesting insights into any agent's policy.

Remote state estimation of large-scale distributed dynamic processes plays an important role in Industry 4.0 applications. In this paper, we focus on the transmission scheduling problem of a remote estimation system. First, we derive some structural properties of the optimal sensor scheduling policy over fading channels. Then, building on these theoretical guidelines, we develop a structure-enhanced deep reinforcement learning (DRL) framework for optimal scheduling of the system to achieve the minimum overall estimation mean-square error (MSE). In particular, we propose a structure-enhanced action selection method, which tends to select actions that obey the policy structure. This explores the action space more effectively and enhances the learning efficiency of DRL agents. Furthermore, we introduce a structure-enhanced loss function to add penalties to actions that do not follow the policy structure. Our numerical experiments illustrate that the proposed structure-enhanced DRL algorithms can save the training time by 50% and reduce the remote estimation MSE by 10% to 25%, when compared to benchmark DRL algorithms. In addition, we show that the derived structural properties exist in a wide range of dynamic scheduling problems that go beyond remote state estimation.

Traffic flow prediction is an important part of smart transportation. The goal is to predict future traffic conditions based on historical data recorded by sensors and the traffic network. As the city continues to build, parts of the transportation network will be added or modified. How to accurately predict expanding and evolving long-term streaming networks is of great significance. To this end, we propose a new simulation-based criterion that considers teaching autonomous agents to mimic sensor patterns, planning their next visit based on the sensor's profile (e.g., traffic, speed, occupancy). The data recorded by the sensor is most accurate when the agent can perfectly simulate the sensor's activity pattern. We propose to formulate the problem as a continuous reinforcement learning task, where the agent is the next flow value predictor, the action is the next time-series flow value in the sensor, and the environment state is a dynamically fused representation of the sensor and transportation network. Actions taken by the agent change the environment, which in turn forces the agent's mode to update, while the agent further explores changes in the dynamic traffic network, which helps the agent predict its next visit more accurately. Therefore, we develop a strategy in which sensors and traffic networks update each other and incorporate temporal context to quantify state representations evolving over time.

Optimally developing an oil and gas field requires predicting future production using a reservoir model, whose key material properties are tuned in a process called history matching. This process of adjusting the key parameters is non-unique and computationally challenging. Typically, the reservoir model is divided into cells that match the geology of the field. The key properties of these cells, such as porosity and permeability, are assigned initially using core sample data, where available. For computational efficiency, the geological model is converted to a reservoir model using upscaling [6, 20, 49] to reduce the number of the cells in the model. Due to the challenges of finding the key properties in each cell, history matching is used to adjust the values of these properties so the model reflects historical production data [19, 28, 9]. History matching is typically done by matching the computed pressure and saturation data (oil, gas and water rates) from the simulation model and comparing it the actual historical data. The difference between the actual data and data generated by the reservoir model is then computed using an objective function that quantifies the mismatch between the two quantities.

Profile extrusion is a continuous production process for manufacturing plastic profiles from molten polymer. Especially interesting is the design of the die, through which the melt is pressed to attain the desired shape. However, due to an inhomogeneous velocity distribution at the die exit or residual stresses inside the extrudate, the final shape of the manufactured part often deviates from the desired one. To avoid these deviations, the shape of the die can be computationally optimized, which has already been investigated in the literature using classical optimization approaches. A new approach in the field of shape optimization is the utilization of Reinforcement Learning (RL) as a learning-based optimization algorithm. RL is based on trial-and-error interactions of an agent with an environment. For each action, the agent is rewarded and informed about the subsequent state of the environment. While not necessarily superior to classical, e.g., gradient-based or evolutionary, optimization algorithms for one single problem, RL techniques are expected to perform especially well when similar optimization tasks are repeated since the agent learns a more general strategy for generating optimal shapes instead of concentrating on just one single problem. In this work, we investigate this approach by applying it to two 2D test cases. The flow-channel geometry can be modified by the RL agent using so-called Free-Form Deformation, a method where the computational mesh is embedded into a transformation spline, which is then manipulated based on the control-point positions. In particular, we investigate the impact of utilizing different agents on the training progress and the potential of wall time saving by utilizing multiple environments during training.

Abstract--This paper proposes a novel centralized training and distributed execution (CTDE)-based multi-agent deep reinforcement learning (MADRL) method for multiple unmanned aerial vehicles (UAVs) control in autonomous mobile access applications. For the purpose, a single neural network is utilized in centralized training for cooperation among multiple agents while maximizing the total quality of service (QoS) in mobile access applications. In order to provide seamless network services in crowded, wild, or extreme areas, which is one of the potential scenarios in 6G networks, the use of unmanned aerial vehicles (UAVs) is widely considered where the UAVs are autonomously operated with deep learning algorithms [1]. In this paper, a multi-agent deep reinforcement learning (MADRL) algorithm is designed and evaluated for autonomous is good enough to utilize the desired performance of multiagent aerial mobile base-station (BS) network coordination cooperation and coordination. In order to neural network, a cost function is required, and the function is achieve our desired goal, one of the promising approaches is designed to maximize the quality of services (QoS) in mobile centralized training and distributed execution (CTDE) where access applications.

Tic Tac Toe is amongst the most well-known games. It has already been shown that it is a biased game, giving more chances to win for the first player leaving only a draw or a loss as possibilities for the opponent, assuming both the players play optimally. Thus on average majority of the games played result in a draw. The majority of the latest research on how to solve a tic tac toe board state employs strategies such as Genetic Algorithms, Neural Networks, Co-Evolution, and Evolutionary Programming. But these approaches deal with a trivial board state of 3X3 and very little research has been done for a generalized algorithm to solve 4X4,5X5,6X6 and many higher states. Even though an algorithm exists which is Min-Max but it takes a lot of time in coming up with an ideal move due to its recursive nature of implementation. A Sample has been created on this link \url{https://bk-tic-tac-toe.herokuapp.com/} to prove this fact. This is the main problem that this study is aimed at solving i.e providing a generalized algorithm(Approximate method, Learning-Based) for higher board states of tic tac toe to make precise moves in a short period. Also, the code changes needed to accommodate higher board states will be nominal. The idea is to pose the tic tac toe game as a well-posed learning problem. The study and its results are promising, giving a high win to draw ratio with each epoch of training. This study could also be encouraging for other researchers to apply the same algorithm to other similar board games like Minesweeper, Chess, and GO for finding efficient strategies and comparing the results.

Warfarin is a widely used anticoagulant, and has a narrow therapeutic range. Dosing of warfarin should be individualized, since slight overdosing or underdosing can have catastrophic or even fatal consequences. Despite much research on warfarin dosing, current dosing protocols do not live up to expectations, especially for patients sensitive to warfarin. We propose a deep reinforcement learning-based dosing model for warfarin. To overcome the issue of relatively small sample sizes in dosing trials, we use a Pharmacokinetic/ Pharmacodynamic (PK/PD) model of warfarin to simulate dose-responses of virtual patients. Applying the proposed algorithm on virtual test patients shows that this model outperforms a set of clinically accepted dosing protocols by a wide margin. We tested the robustness of our dosing protocol on a second PK/PD model and showed that its performance is comparable to the set of baseline protocols.

Despite its popularity in the reinforcement learning community, a provably convergent policy gradient method for continuous space-time control problems with nonlinear state dynamics has been elusive. This paper proposes proximal gradient algorithms for feedback controls of finite-time horizon stochastic control problems. The state dynamics are nonlinear diffusions with control-affine drift, and the cost functions are nonconvex in the state and nonsmooth in the control. The system noise can degenerate, which allows for deterministic control problems as special cases. We prove under suitable conditions that the algorithm converges linearly to a stationary point of the control problem, and is stable with respect to policy updates by approximate gradient steps. The convergence result justifies the recent reinforcement learning heuristics that adding entropy regularization or a fictitious discount factor to the optimization objective accelerates the convergence of policy gradient methods.