Empirical design in reinforcement learning is no small task. Running good experiments requires attention to detail and at times significant computational resources. While compute resources available per dollar have continued to grow rapidly, so have the scale of typical experiments in reinforcement learning. It is now common to benchmark agents with millions of parameters against dozens of tasks, each using the equivalent of 30 days of experience. The scale of these experiments often conflict with the need for proper statistical evidence, especially when comparing algorithms. Recent studies have highlighted how popular algorithms are sensitive to hyper-parameter settings and implementation details, and that common empirical practice leads to weak statistical evidence (Machado et al., 2018; Henderson et al., 2018). Here we take this one step further. This manuscript represents both a call to action, and a comprehensive resource for how to do good experiments in reinforcement learning. In particular, we cover: the statistical assumptions underlying common performance measures, how to properly characterize performance variation and stability, hypothesis testing, special considerations for comparing multiple agents, baseline and illustrative example construction, and how to deal with hyper-parameters and experimenter bias. Throughout we highlight common mistakes found in the literature and the statistical consequences of those in example experiments. The objective of this document is to provide answers on how we can use our unprecedented compute to do good science in reinforcement learning, as well as stay alert to potential pitfalls in our empirical design.

Agricultural irrigation is a significant contributor to freshwater consumption. However, the current irrigation systems used in the field are not efficient. They rely mainly on soil moisture sensors and the experience of growers, but do not account for future soil moisture loss. Predicting soil moisture loss is challenging because it is influenced by numerous factors, including soil texture, weather conditions, and plant characteristics. This paper proposes a solution to improve irrigation efficiency, which is called DRLIC. DRLIC is a sophisticated irrigation system that uses deep reinforcement learning (DRL) to optimize its performance. The system employs a neural network, known as the DRL control agent, which learns an optimal control policy that considers both the current soil moisture measurement and the future soil moisture loss. We introduce an irrigation reward function that enables our control agent to learn from previous experiences. However, there may be instances where the output of our DRL control agent is unsafe, such as irrigating too much or too little water. To avoid damaging the health of the plants, we implement a safety mechanism that employs a soil moisture predictor to estimate the performance of each action. If the predicted outcome is deemed unsafe, we perform a relatively-conservative action instead. To demonstrate the real-world application of our approach, we developed an irrigation system that comprises sprinklers, sensing and control nodes, and a wireless network. We evaluate the performance of DRLIC by deploying it in a testbed consisting of six almond trees. During a 15-day in-field experiment, we compared the water consumption of DRLIC with a widely-used irrigation scheme. Our results indicate that DRLIC outperformed the traditional irrigation method by achieving a water savings of up to 9.52%.

Reinforcement learning (RL) is a branch of machine learning that has been used in a variety of applications such as robotics, game playing, and autonomous systems. In recent years, there has been growing interest in applying RL to quantitative trading, where the goal is to make profitable trades in financial markets. This paper explores the use of RL in quantitative trading and presents a case study of a RL-based trading algorithm. The results show that RL can be a powerful tool for quantitative trading, and that it has the potential to outperform traditional trading algorithms. The use of reinforcement learning in quantitative trading represents a promising area of research that can potentially lead to the development of more sophisticated and effective trading systems. Future work could explore the use of alternative reinforcement learning algorithms, incorporate additional data sources, and test the system on different asset classes. Overall, our research demonstrates the potential of using reinforcement learning in quantitative trading and highlights the importance of continued research and development in this area. By developing more sophisticated and effective trading systems, we can potentially improve the efficiency of financial markets and generate greater returns for investors.

Autonomous cyber agents may be developed by applying reinforcement and deep reinforcement learning (RL/DRL), where agents are trained in a representative environment. The training environment must simulate with high-fidelity the network Cyber Operations (CyOp) that the agent aims to explore. Given the complexity of net-work CyOps, a good simulator is difficult to achieve. This work presents a systematic solution to automatically generate a high-fidelity simulator in the Cyber Gym for Intelligent Learning (CyGIL). Through representation learning and continuous learning, CyGIL provides a unified CyOp training environment where an emulated CyGIL-E automatically generates a simulated CyGIL-S. The simulator generation is integrated with the agent training process to further reduce the required agent training time. The agent trained in CyGIL-S is transferrable directly to CyGIL-E showing full transferability to the emulated "real" network. Experimental results are presented to demonstrate the CyGIL training performance. Enabling offline RL, the CyGIL solution presents a promising direction towards sim-to-real for leveraging RL agents in real-world cyber networks.

In this paper, we present a brief survey of Reinforcement Learning (RL), with particular emphasis on Stochastic Approximation (SA) as a unifying theme. The scope of the paper includes Markov Reward Processes, Markov Decision Processes, Stochastic Approximation methods, and widely used algorithms such as Temporal Difference Learning and Q-learning. Reinforcement Learning is a vast subject, and this brief survey can barely do justice to the topic. There are several excellent texts on RL, such as [4, 27, 34, 33]. The dynamics of the Stochastic Approximation (SA) algorithm are analyzed in [25, 22, 3, 23, 2, 9, 10]. The interested reader may consult those sources for more information. In this survey, we use the phrase "reinforcement learning" to refer to decision-making with uncertain models, and in addition, current actions alter the future behavior of the system. Therefore, if the same action is taken at a future time, the consequences might not be the same.

In this paper, we propose a novel framework for tactile-based dexterous manipulation learning with a blind anthropomorphic robotic hand, i.e. without visual sensing. First, object-related states were extracted from the raw tactile signals by a graph-based perception model - TacGNN. The resulting tactile features were then utilized in the policy learning of an in-hand manipulation task in the second stage. This method was examined by a Baoding ball task - simultaneously manipulating two spheres around each other by 180 degrees in hand. We conducted experiments on object states prediction and in-hand manipulation using a reinforcement learning algorithm (PPO). Results show that TacGNN is effective in predicting object-related states during manipulation by decreasing the RMSE of prediction to 0.096cm comparing to other methods, such as MLP, CNN, and GCN. Finally, the robot hand could finish an in-hand manipulation task solely relying on the robotic own perception - tactile sensing and proprioception. In addition, our methods are tested on three tasks with different difficulty levels and transferred to the real robot without further training.

Most reinforcement learning algorithms are based on a key assumption that Markov decision processes (MDPs) are stationary. However, non-stationary MDPs with dynamic action space are omnipresent in real-world scenarios. Yet problems of dynamic action space reinforcement learning have been studied by many previous works, how to choose valuable actions from new and unseen actions to improve learning efficiency remains unaddressed. To tackle this problem, we propose an intelligent Action Pick-up (AP) algorithm to autonomously choose valuable actions that are most likely to boost performance from a set of new actions. In this paper, we first theoretically analyze and find that a prior optimal policy plays an important role in action pick-up by providing useful knowledge and experience. Then, we design two different AP methods: frequency-based global method and state clustering-based local method, based on the prior optimal policy. Finally, we evaluate the AP on two simulated but challenging environments where action spaces vary over time. Experimental results demonstrate that our proposed AP has advantages over baselines in learning efficiency.

There is a recent surge in interest for imitation learning, with large human video-game and robotic manipulation datasets being used to train agents on very complex tasks. While deep neuroevolution has recently been shown to match the performance of gradient-based techniques on various reinforcement learning problems, the application of deep neuroevolution techniques to imitation learning remains relatively unexplored. In this work, we propose to explore whether deep neuroevolution can be used for behaviour imitation on popular simulation environments. We introduce a simple co-evolutionary adversarial generation framework, and evaluate its capabilities by evolving standard deep recurrent networks to imitate state-of-the-art pre-trained agents on 8 OpenAI Gym state-based control tasks. Across all tasks, we find the final elite actor agents capable of achieving scores as high as those obtained by the pre-trained agents, all the while closely following their score trajectories. Our results suggest that neuroevolution could be a valuable addition to deep learning techniques to produce accurate emulation of behavioural agents. We believe that the generality and simplicity of our approach opens avenues for imitating increasingly complex behaviours in increasingly complex settings, e.g. human behaviour in real-world settings. We provide our source code, model checkpoints and results at github.com/MaximilienLC/gane.

Hardening cyber physical assets is both crucial and labor-intensive. Recently, Machine Learning (ML) in general and Reinforcement Learning RL) more specifically has shown great promise to automate tasks that otherwise would require significant human insight/intelligence. The development of autonomous RL agents requires a suitable training environment that allows us to quickly evaluate various alternatives, in particular how to arrange training scenarios that pit attackers and defenders against each other. CyberBattleSim is a training environment that supports the training of red agents, i.e., attackers. We added the capability to train blue agents, i.e., defenders. The paper describes our changes and reports on the results we obtained when training blue agents, either in isolation or jointly with red agents. Our results show that training a blue agent does lead to stronger defenses against attacks. In particular, training a blue agent jointly with a red agent increases the blue agent's capability to thwart sophisticated red agents.

This work aims to enable autonomous agents for network cyber operations (CyOps) by applying reinforcement and deep reinforcement learning (RL/DRL). The required RL training environment is particularly challenging, as it must balance the need for high-fidelity, best achieved through real network emulation, with the need for running large numbers of training episodes, best achieved using simulation. A unified training environment, namely the Cyber Gym for Intelligent Learning (CyGIL) is developed where an emulated CyGIL-E automatically generates a simulated CyGIL-S. From preliminary experimental results, CyGIL-S is capable to train agents in minutes compared with the days required in CyGIL-E. The agents trained in CyGIL-S are transferrable directly to CyGIL-E showing full decision proficiency in the emulated "real" network. Enabling offline RL, the CyGIL solution presents a promising direction towards sim-to-real for leveraging RL agents in real-world cyber networks.