We consider an extended notion of reinforcement learning in which the environment can simulate the agent and base its outputs on the agent's hypothetical behavior. Since good performance usually requires paying attention to whatever things the environment's outputs are based on, we argue that for an agent to achieve on-average good performance across many such extended environments, it is necessary for the agent to self-reflect. Thus, an agent's self-reflection ability can be numerically estimated by running the agent through a battery of extended environments. We are simultaneously releasing an open-source library of extended environments to serve as proof-of-concept of this technique. As the library is first-of-kind, we have avoided the difficult problem of optimizing it. Instead we have chosen environments with interesting properties. Some seem paradoxical, some lead to interesting thought experiments, some are even suggestive of how self-reflection might have evolved in nature. We give examples and introduce a simple transformation which experimentally seems to increase self-reflection.

When learning common skills like driving, beginners usually have domain experts standing by to ensure the safety of the learning process. We formulate such learning scheme under the Expert-in-the-loop Reinforcement Learning where a guardian is introduced to safeguard the exploration of the learning agent. While allowing the sufficient exploration in the uncertain environment, the guardian intervenes under dangerous situations and demonstrates the correct actions to avoid potential accidents. Thus ERL enables both exploration and expert's partial demonstration as two training sources. Following such a setting, we develop a novel Expert Guided Policy Optimization (EGPO) method which integrates the guardian in the loop of reinforcement learning. The guardian is composed of an expert policy to generate demonstration and a switch function to decide when to intervene. Particularly, a constrained optimization technique is used to tackle the trivial solution that the agent deliberately behaves dangerously to deceive the expert into taking over. Offline RL technique is further used to learn from the partial demonstration generated by the expert. Safe driving experiments show that our method achieves superior training and test-time safety, outperforms baselines with a substantial margin in sample efficiency, and preserves the generalizabiliy to unseen environments in test-time. Demo video and source code are available at: https://decisionforce.github.io/EGPO/

Low earth orbit (LEO) satellite-assisted communications have been considered as one of key elements in beyond 5G systems to provide wide coverage and cost-efficient data services. Such dynamic space-terrestrial topologies impose exponential increase in the degrees of freedom in network management. In this paper, we address two practical issues for an over-loaded LEO-terrestrial system. The first challenge is how to efficiently schedule resources to serve the massive number of connected users, such that more data and users can be delivered/served. The second challenge is how to make the algorithmic solution more resilient in adapting to dynamic wireless environments.To address them, we first propose an iterative suboptimal algorithm to provide an offline benchmark. To adapt to unforeseen variations, we propose an enhanced meta-critic learning algorithm (EMCL), where a hybrid neural network for parameterization and the Wolpertinger policy for action mapping are designed in EMCL. The results demonstrate EMCL's effectiveness and fast-response capabilities in over-loaded systems and in adapting to dynamic environments compare to previous actor-critic and meta-learning methods.

We consider the problem of using expert data with unobserved confounders for imitation and reinforcement learning. We begin by defining the problem of learning from confounded expert data in a contextual MDP setup. We analyze the limitations of learning from such data with and without external reward, and propose an adjustment of standard imitation learning algorithms to fit this setup. We then discuss the problem of distribution shift between the expert data and the online environment when the data is only partially observable. We prove possibility and impossibility results for imitation learning under arbitrary distribution shift of the missing covariates. When additional external reward is provided, we propose a sampling procedure that addresses the unknown shift and prove convergence to an optimal solution. Finally, we validate our claims empirically on challenging assistive healthcare and recommender system simulation tasks.

In this work, a novel idea is presented for combinatorial optimization problems, a hybrid network, which results in a superior outcome. We applied this method to graph pointer networks [1], expanding its capabilities to a higher level. We proposed a hybrid pointer network (HPN) to solve the travelling salesman problem trained by reinforcement learning. Furthermore, HPN builds upon graph pointer networks which is an extension of pointer networks with an additional graph embedding layer. HPN outperforms the graph pointer network in solution quality due to the hybrid encoder, which provides our model with a verity encoding type, allowing our model to converge to a better policy. Our network significantly outperforms the original graph pointer network for small and large-scale problems increasing its performance for TSP50 from 5.959 to 5.706 without utilizing 2opt, Pointer networks, Attention model, and a wide range of models, producing results comparable to highly tuned and specialized algorithms. We make our data, models, and code publicly available [2].

Communications standards are designed via committees of humans holding repeated meetings over months or even years until consensus is achieved. This includes decisions regarding the modulation and coding schemes to be supported over an air interface. We propose a way to "automate" the selection of the set of modulation and coding schemes to be supported over a given air interface and thereby streamline both the standards design process and the ease of extending the standard to support new modulation schemes applicable to new higher-level applications and services. Our scheme involves machine learning, whereby a constructor entity submits proposals to an evaluator entity, which returns a score for the proposal. The constructor employs reinforcement learning to iterate on its submitted proposals until a score is achieved that was previously agreed upon by both constructor and evaluator to be indicative of satisfying the required design criteria (including performance metrics for transmissions over the interface).

As a major value function evaluation method, temporal difference (TD) learning (Sutton, 1988; Dayan, 1992) has been widely used in various planning problems in reinforcement learning. Although TD learning performs successfully in the on-policy settings, where an agent can interact with environments under the target policy, it can perform poorly or even diverge under the off-policy settings when the agent only has access to data sampled by a behavior policy (Baird, 1995; Tsitsiklis and Van Roy, 1997; Mahmood et al., 2015). To address such an issue, the gradient temporal-difference (GTD) (Sutton et al., 2008) and least-squares temporal difference (LSTD) (Yu, 2010) algorithms have been proposed, which have been shown to converge in the off-policy settings. However, since GTD and LSTD consider an objective function based on the behavior policy, their converging points can be largely biased from the true value function due to the distribution mismatch between the target and behavior policies, even when the express power of the function approximation class is arbitrarily large (Kolter, 2011). In order to provide a more accurate evaluation, Sutton et al. (2016) proposed the emphatic temporal difference (ETD) algorithm, which introduces the follow-on trace to address the distribution mismatch issue. The stability of ETD was then shown in Sutton et al. (2016); Mahmood et al. (2015), and the asymptotic convergence guarantee for ETD was established in Yu (2015), it has also achieved great success in many tasks (Ghiassian et al., 2016; Ni, 2021).

Reading to act is a prevalent but challenging task which requires the ability to reason from a concise instruction. However, previous works face the semantic mismatch between the low-level actions and the high-level language descriptions and require the human-designed curriculum to work properly. In this paper, we present a Feudal Reinforcement Learning (FRL) model consisting of a manager agent and a worker agent. The manager agent is a multi-hop plan generator dealing with high-level abstract information and generating a series of sub-goals in a backward manner. The worker agent deals with the low-level perceptions and actions to achieve the sub-goals one by one. In comparison, our FRL model effectively alleviate the mismatching between text-level inference and low-level perceptions and actions; and is general to various forms of environments, instructions and manuals; and our multi-hop plan generator can significantly boost for challenging tasks where multi-step reasoning form the texts is critical to resolve the instructed goals. We showcase our approach achieves competitive performance on two challenging tasks, Read to Fight Monsters (RTFM) and Messenger, without human-designed curriculum learning. Recently, there are increasing interests in building reinforcement learning (RL) agents that interact with humans via natural language, such as follow natural language instructions and complete goals specified in natural language. The successes of these studies will boost the user experience in a wide range of real-world applications, such as visual language navigation (Anderson et al., 2018; Wang et al., 2019b), interactive games (Gray et al., 2019), robot control (Tellex et al., 2020), goal-oriented dialog systems and other personal assistant applications (Dhingra et al., 2017). In order to generalize to real-world use cases, the research of RL with language instructions faces various kinds of complexity. One critical demand of these use cases is that humans tend to give concise instructions, which specify the goals they hope to achieve, instead of providing complete information for the intermediate steps.

Increasing amounts of distributed generation in distribution networks can provide both challenges and opportunities for voltage regulation across the network. Intelligent control of smart inverters and other smart building energy management systems can be leveraged to alleviate these issues. GridLearn is a multiagent reinforcement learning platform that incorporates both building energy models and power flow models to achieve grid level goals, by controlling behind-the-meter resources. This study demonstrates how multi-agent reinforcement learning can preserve building owner privacy and comfort while pursuing grid-level objectives. Building upon the CityLearn framework which considers RL for building-level goals, this work expands the framework to a network setting where grid-level goals are additionally considered. As a case study, we consider voltage regulation on the IEEE-33 bus network using controllable building loads, energy storage, and smart inverters. The results show that the RL agents nominally reduce instances of undervoltages and reduce instances of overvoltages by 34%.

Reinforcement Learning (RL) can be considered as a sequence modeling task, i.e., given a sequence of past state-action-reward experiences, a model autoregressively predicts a sequence of future actions. Recently, Transformers have been successfully adopted to model this problem. In this work, we propose State-Action-Reward Transformer (StARformer), which explicitly models local causal relations to help improve action prediction in long sequences. A sequence of such local representations combined with state representations, is then used to make action predictions over a long time span. Our experiments show that StARformer outperforms the state-of-the-art Transformer-based method on Atari (image) and Gym (state vector) benchmarks, in both offline-RL and imitation learning settings. StARformer is also more compliant with longer sequences of inputs compared to the baseline. Our code is available at https://github.com/ Reinforcement Learning (RL) naturally comes with sequential data: an agent observes a state from the environment, takes an action, observes the next state and receives a reward from the environment.