Past research has proposed numerous hardware prefetching techniques, most of which rely on exploiting one specific type of program context information (e.g., program counter, cacheline address) to predict future memory accesses. These techniques either completely neglect a prefetcher's undesirable effects (e.g., memory bandwidth usage) on the overall system, or incorporate system-level feedback as an afterthought to a system-unaware prefetch algorithm. We show that prior prefetchers often lose their performance benefit over a wide range of workloads and system configurations due to their inherent inability to take multiple different types of program context and system-level feedback information into account while prefetching. In this paper, we make a case for designing a holistic prefetch algorithm that learns to prefetch using multiple different types of program context and system-level feedback information inherent to its design. To this end, we propose Pythia, which formulates the prefetcher as a reinforcement learning agent. For every demand request, Pythia observes multiple different types of program context information to make a prefetch decision. For every prefetch decision, Pythia receives a numerical reward that evaluates prefetch quality under the current memory bandwidth usage. Pythia uses this reward to reinforce the correlation between program context information and prefetch decision to generate highly accurate, timely, and system-aware prefetch requests in the future. Our extensive evaluations using simulation and hardware synthesis show that Pythia outperforms multiple state-of-the-art prefetchers over a wide range of workloads and system configurations, while incurring only 1.03% area overhead over a desktop-class processor and no software changes in workloads. The source code of Pythia can be freely downloaded from https://github.com/CMU-SAFARI/Pythia.

We propose a framework for the design of feedback controllers that combines the optimization-driven and model-free advantages of deep reinforcement learning with the stability guarantees provided by using the Youla-Kucera parameterization to define the search domain. Recent advances in behavioral systems allow us to construct a data-driven internal model; this enables an alternative realization of the Youla-Kucera parameterization based entirely on input-output exploration data. Using a neural network to express a parameterized set of nonlinear stable operators enables seamless integration with standard deep learning libraries. We demonstrate the approach on a realistic simulation of a two-tank system.

Offline reinforcement learning (RL) methods can generally be categorized into two types: RL-based and Imitation-based. RL-based methods could in principle enjoy out-of-distribution generalization but suffer from erroneous off-policy evaluation. Imitation-based methods avoid off-policy evaluation but are too conservative to surpass the dataset. In this study, we propose an alternative approach, inheriting the training stability of imitation-style methods while still allowing logical out-of-distribution generalization. We decompose the conventional reward-maximizing policy in offline RL into a guide-policy and an execute-policy. During training, the guide-poicy and execute-policy are learned using only data from the dataset, in a supervised and decoupled manner. During evaluation, the guide-policy guides the execute-policy by telling where it should go so that the reward can be maximized, serving as the \textit{Prophet}. By doing so, our algorithm allows \textit{state-compositionality} from the dataset, rather than \textit{action-compositionality} conducted in prior imitation-style methods. We dumb this new approach Policy-guided Offline RL (\texttt{POR}). \texttt{POR} demonstrates the state-of-the-art performance on D4RL, a standard benchmark for offline RL. We also highlight the benefits of \texttt{POR} in terms of improving with supplementary suboptimal data and easily adapting to new tasks by only changing the guide-poicy.

Despite their prevalence in eHealth applications for behavior change, persuasive messages tend to have small effects on behavior. Conditions or states (e.g., confidence, knowledge, motivation) and characteristics (e.g., gender, age, personality) of persuadees are two promising components for more effective algorithms for choosing persuasive messages. However, it is not yet sufficiently clear how well considering these components allows one to predict behavior after persuasive attempts, especially in the long run. Since collecting data for many algorithm components is costly and places a burden on users, a better understanding of the impact of individual components in practice is welcome. This can help to make an informed decision on which components to use. We thus conducted a longitudinal study in which a virtual coach persuaded 671 daily smokers to do preparatory activities for quitting smoking and becoming more physically active, such as envisioning one's desired future self. Based on the collected data, we designed a Reinforcement Learning (RL)-approach that considers current and future states to maximize the effort people spend on their activities. Using this RL-approach, we found, based on leave-one-out cross-validation, that considering states helps to predict both behavior and future states. User characteristics and especially involvement in the activities, on the other hand, only help to predict behavior if used in combination with states rather than alone. We see these results as supporting the use of states and involvement in persuasion algorithms. Our dataset is available online.

In reinforcement learning, exploration refers to the agent taking actions according to a behavior policy in order to traverse a typically discrete state space and collect rewards. While exploring the state space, the agent uses an update rule to estimate, based on the rewards collected, the Q-values (i.e., state-action values) from one iteration to the next. If the Q-values converge to their optimums, an optimal policy can then be obtained. For a class of reinforcement learning problems, such convergence is guaranteed under the Robbins-Monro conditions [47]. A requirement for satisfying the Robbins-Monro conditions is that every state-action pair must have a non-zero probability of being visited by the agent -- also known as persistent exploration. If we consider the agent taking an action (when it is at a state) as'generating' a symbol denoting that action, the sequences of actions thus generated by the agent as it traverses through the states represent the behavior of the agent. For an episodic learning process, the behavior of the agent consists of all possible action sequences from the initial state to the set of goal states. We refer to such a process as an unconstrained learning process, and the associated optimal Q-values as the intrinsic optimums.

Single occupancy vehicles are the most attractive transportation alternative for many commuters, leading to increased traffic congestion and air pollution. Advancements in information technologies create opportunities for smart solutions that incentivize ridesharing and mode shift to higher occupancy vehicles (HOVs) to achieve the car lighter vision of cities. In this study, we present HumanLight, a novel decentralized adaptive traffic signal control algorithm designed to optimize people throughput at intersections. Our proposed controller is founded on reinforcement learning with the reward function embedding the transportation-inspired concept of pressure at the person-level. By rewarding HOV commuters with travel time savings for their efforts to merge into a single ride, HumanLight achieves equitable allocation of green times. Apart from adopting FRAP, a state-of-the-art (SOTA) base model, HumanLight introduces the concept of active vehicles, loosely defined as vehicles in proximity to the intersection within the action interval window. The proposed algorithm showcases significant headroom and scalability in different network configurations considering multimodal vehicle splits at various scenarios of HOV adoption. Improvements in person delays and queues range from 15% to over 55% compared to vehicle-level SOTA controllers. We quantify the impact of incorporating active vehicles in the formulation of our RL model for different network structures. HumanLight also enables regulation of the aggressiveness of the HOV prioritization. The impact of parameter setting on the generated phase profile is investigated as a key component of acyclic signal controllers affecting pedestrian waiting times. HumanLight's scalable, decentralized design can reshape the resolution of traffic management to be more human-centric and empower policies that incentivize ridesharing and public transit systems.

In the traditional vehicular network, computing tasks generated by the vehicles are usually uploaded to the cloud for processing. However, since task offloading toward the cloud will cause a large delay, vehicular edge computing (VEC) is introduced to avoid such a problem and improve the whole system performance, where a roadside unit (RSU) with certain computing capability is used to process the data of vehicles as an edge entity. Owing to the privacy and security issues, vehicles are reluctant to upload local data directly to the RSU, and thus federated learning (FL) becomes a promising technology for some machine learning tasks in VEC, where vehicles only need to upload the local model hyperparameters instead of transferring their local data to the nearby RSU. Furthermore, as vehicles have different local training time due to various sizes of local data and their different computing capabilities, asynchronous federated learning (AFL) is employed to facilitate the RSU to update the global model immediately after receiving a local model to reduce the aggregation delay. However, in AFL of VEC, different vehicles may have different impact on the global model updating because of their various local training delay, transmission delay and local data sizes. Also, if there are bad nodes among the vehicles, it will affect the global aggregation quality at the RSU. To solve the above problem, we shall propose a deep reinforcement learning (DRL) based vehicle selection scheme to improve the accuracy of the global model in AFL of vehicular network. In the scheme, we present the model including the state, action and reward in the DRL based to the specific problem. Simulation results demonstrate our scheme can effectively remove the bad nodes and improve the aggregation accuracy of the global model.

Wind energy has been rapidly gaining popularity as a means for combating climate change. However, the variable nature of wind generation can undermine system reliability and lead to wind curtailment, causing substantial economic losses to wind power producers. Battery energy storage systems (BESS) that serve as onsite backup sources are among the solutions to mitigate wind curtailment. However, such an auxiliary role of the BESS might severely weaken its economic viability. This paper addresses the issue by proposing joint wind curtailment reduction and energy arbitrage for the BESS. We decouple the market participation of the co-located wind-battery system and develop a joint-bidding framework for the wind farm and BESS. It is challenging to optimize the joint-bidding because of the stochasticity of energy prices and wind generation. Therefore, we leverage deep reinforcement learning to maximize the overall revenue from the spot market while unlocking the BESS's potential in concurrently reducing wind curtailment and conducting energy arbitrage. We validate the proposed strategy using realistic wind farm data and demonstrate that our joint-bidding strategy responds better to wind curtailment and generates higher revenues than the optimization-based benchmark. Our simulations also reveal that the extra wind generation used to be curtailed can be an effective power source to charge the BESS, resulting in additional financial returns.

Complex reasoning problems contain states that vary in the computational cost required to determine a good action plan. Taking advantage of this property, we propose Adaptive Subgoal Search (AdaSubS), a search method that adaptively adjusts the planning horizon. To this end, AdaSubS generates diverse sets of subgoals at different distances. A verification mechanism is employed to filter out unreachable subgoals swiftly, allowing to focus on feasible further subgoals. In this way, AdaSubS benefits from the efficiency of planning with longer subgoals and the fine control with the shorter ones, and thus scales well to difficult planning problems. We show that AdaSubS significantly surpasses hierarchical planning algorithms on three complex reasoning tasks: Sokoban, the Rubik's Cube, and inequality proving benchmark INT.

Autonomous cyber and cyber-physical systems need to perform decision-making, learning, and control in unknown environments. Such decision-making can be sensitive to multiple factors, including modeling errors, changes in costs, and impacts of events in the tails of probability distributions. Although multi-agent reinforcement learning (MARL) provides a framework for learning behaviors through repeated interactions with the environment by minimizing an average cost, it will not be adequate to overcome the above challenges. In this paper, we develop a distributed MARL approach to solve decision-making problems in unknown environments by learning risk-aware actions. We use the conditional value-at-risk (CVaR) to characterize the cost function that is being minimized, and define a Bellman operator to characterize the value function associated to a given state-action pair. We prove that this operator satisfies a contraction property, and that it converges to the optimal value function. We then propose a distributed MARL algorithm called the CVaR QD-Learning algorithm, and establish that value functions of individual agents reaches consensus. We identify several challenges that arise in the implementation of the CVaR QD-Learning algorithm, and present solutions to overcome these. We evaluate the CVaR QD-Learning algorithm through simulations, and demonstrate the effect of a risk parameter on value functions at consensus.