Deep Reinforcement Learning (RL) has been demonstrated to yield capable agents and control policies in several domains but is commonly plagued by prohibitively long training times. Additionally, in the case of continuous control problems, the applicability of learned policies on real-world embedded devices is limited due to the lack of real-time guarantees and portability of existing deep learning libraries. To address these challenges, we present RLtools, a dependency-free, header-only, pure C++ library for deep supervised and reinforcement learning. Leveraging the template meta-programming capabilities of recent C++ standards, we provide composable components that can be tightly integrated by the compiler. Its novel architecture allows RLtools to be used seamlessly on a heterogeneous set of platforms, from HPC clusters over workstations and laptops to smartphones, smartwatches, and microcontrollers. Specifically, due to the tight integration of the RL algorithms with simulation environments, RLtools can solve popular RL problems like the Pendulum-v1 swing-up about 7 to 15 times faster in terms of wall-clock training time compared to other popular RL frameworks when using TD3. We also provide a low-overhead and parallelized interface to the MuJoCo simulator, showing that our PPO implementation achieves state of the art returns in the Ant-v4 environment while being 25%-30% faster in terms of wall-clock training time. Finally, we also benchmark the policy inference on a diverse set of microcontrollers and show that in most cases our optimized inference implementation is much faster than even the manufacturer's DSP libraries. To the best of our knowledge, RLtools enables the first-ever demonstration of training a deep RL algorithm directly on a microcontroller, giving rise to the field of TinyRL. The source code is available through our project page at https://rl.tools.

Reinforcement Learning views the maximization of rewards and avoidance of punishments as central to explaining goal-directed behavior. However, over a life, organisms will need to learn about many different aspects of the world's structure: the states of the world and state-vector transition dynamics. The number of combinations of states grows exponentially as an agent incorporates new knowledge, and there is no obvious weighted combination of pre-existing rewards or costs defined for a given combination of states, as such a weighting would need to encode information about good and bad combinations prior to an agent's experience in the world. Therefore, we must develop more naturalistic accounts of behavior and motivation in large state-spaces. We show that it is possible to use only the intrinsic motivation metric of empowerment, which measures the agent's capacity to realize many possible futures under a transition operator. We propose to scale empowerment to hierarchical state-spaces by using Operator Bellman Equations. These equations produce state-time feasibility functions, which are compositional hierarchical state-time transition operators that map an initial state and time when an agent begins a policy to the final states and times of completing a goal. Because these functions are hierarchical operators we can define hierarchical empowerment measures on them. An agent can then optimize plans to distant states and times to maximize its hierarchical empowerment-gain, allowing it to discover goals that bring about a more favorable coupling of its internal structure (physiological states) to its external environment (world structure & spatial state). Life-long agents could therefore be primarily animated by principles of compositionality and empowerment, exhibiting self-concern for the growth & maintenance of their own structural integrity without recourse to reward-maximization.

We introduce and study constrained Markov Decision Processes (cMDPs) with anytime constraints. An anytime constraint requires the agent to never violate its budget at any point in time, almost surely. Although Markovian policies are no longer sufficient, we show that there exist optimal deterministic policies augmented with cumulative costs. In fact, we present a fixed-parameter tractable reduction from anytime-constrained cMDPs to unconstrained MDPs. Our reduction yields planning and learning algorithms that are time and sample-efficient for tabular cMDPs so long as the precision of the costs is logarithmic in the size of the cMDP. However, we also show that computing non-trivial approximately optimal policies is NP-hard in general. To circumvent this bottleneck, we design provable approximation algorithms that efficiently compute or learn an arbitrarily accurate approximately feasible policy with optimal value so long as the maximum supported cost is bounded by a polynomial in the cMDP or the absolute budget. Given our hardness results, our approximation guarantees are the best possible under worst-case analysis.

We investigate the problem of corruption robustness in offline reinforcement learning (RL) with general function approximation, where an adversary can corrupt each sample in the offline dataset, and the corruption level $\zeta\geq0$ quantifies the cumulative corruption amount over $n$ episodes and $H$ steps. Our goal is to find a policy that is robust to such corruption and minimizes the suboptimality gap with respect to the optimal policy for the uncorrupted Markov decision processes (MDPs). Drawing inspiration from the uncertainty-weighting technique from the robust online RL setting \citep{he2022nearly,ye2022corruptionrobust}, we design a new uncertainty weight iteration procedure to efficiently compute on batched samples and propose a corruption-robust algorithm for offline RL. Notably, under the assumption of single policy coverage and the knowledge of $\zeta$, our proposed algorithm achieves a suboptimality bound that is worsened by an additive factor of $\mathcal O(\zeta \cdot (\text{CC}(\lambda,\hat{\mathcal F},\mathcal Z_n^H))^{1/2} (C(\hat{\mathcal F},\mu))^{-1/2} n^{-1})$ due to the corruption. Here $\text{CC}(\lambda,\hat{\mathcal F},\mathcal Z_n^H)$ is the coverage coefficient that depends on the regularization parameter $\lambda$, the confidence set $\hat{\mathcal F}$, and the dataset $\mathcal Z_n^H$, and $C(\hat{\mathcal F},\mu)$ is a coefficient that depends on $\hat{\mathcal F}$ and the underlying data distribution $\mu$. When specialized to linear MDPs, the corruption-dependent error term reduces to $\mathcal O(\zeta d n^{-1})$ with $d$ being the dimension of the feature map, which matches the existing lower bound for corrupted linear MDPs. This suggests that our analysis is tight in terms of the corruption-dependent term.

Future sequence represents the outcome after executing the action into the environment (i.e. the trajectory onwards). When driven by the information-theoretic concept of mutual information, it seeks maximally informative consequences. Explicit outcomes may vary across state, return, or trajectory serving different purposes such as credit assignment or imitation learning. However, the inherent nature of incorporating intrinsic motivation with reward maximization is often neglected. In this work, we propose a policy iteration scheme that seamlessly incorporates the mutual information, ensuring convergence to the optimal policy. Concurrently, a variational approach is introduced, which jointly learns the necessary quantity for estimating the mutual information and the dynamics model, providing a general framework for incorporating different forms of outcomes of interest. While we mainly focus on theoretical analysis, our approach opens the possibilities of leveraging intrinsic control with model learning to enhance sample efficiency and incorporate uncertainty of the environment into decision-making.

Hybrid RL is the setting where an RL agent has access to both offline data and online data by interacting with the real-world environment. In this work, we propose a new hybrid RL algorithm that combines an on-policy actor-critic method with offline data. On-policy methods such as policy gradient and natural policy gradient (NPG) have shown to be more robust to model misspecification, though sometimes it may not be as sample efficient as methods that rely on off-policy learning. On the other hand, offline methods that depend on off-policy training often require strong assumptions in theory and are less stable to train in practice. Our new approach integrates a procedure of off-policy training on the offline data into an on-policy NPG framework. We show that our approach, in theory, can obtain a best-of-both-worlds type of result -- it achieves the state-of-art theoretical guarantees of offline RL when offline RL-specific assumptions hold, while at the same time maintaining the theoretical guarantees of on-policy NPG regardless of the offline RL assumptions' validity. Experimentally, in challenging rich-observation environments, we show that our approach outperforms a state-of-the-art hybrid RL baseline which only relies on off-policy policy optimization, demonstrating the empirical benefit of combining on-policy and off-policy learning. Our code is publicly available at https://github.com/YifeiZhou02/HNPG.

We present RoboGen, a generative robotic agent that automatically learns diverse robotic skills at scale via generative simulation. RoboGen leverages the latest advancements in foundation and generative models. Instead of directly using or adapting these models to produce policies or low-level actions, we advocate for a generative scheme, which uses these models to automatically generate diversified tasks, scenes, and training supervisions, thereby scaling up robotic skill learning with minimal human supervision. Our approach equips a robotic agent with a self-guided propose-generate-learn cycle: the agent first proposes interesting tasks and skills to develop, and then generates corresponding simulation environments by populating pertinent objects and assets with proper spatial configurations. Afterwards, the agent decomposes the proposed high-level task into sub-tasks, selects the optimal learning approach (reinforcement learning, motion planning, or trajectory optimization), generates required training supervision, and then learns policies to acquire the proposed skill. Our work attempts to extract the extensive and versatile knowledge embedded in large-scale models and transfer them to the field of robotics. Our fully generative pipeline can be queried repeatedly, producing an endless stream of skill demonstrations associated with diverse tasks and environments.

In this work, we study the low-rank MDPs with adversarially changed losses in the full-information feedback setting. In particular, the unknown transition probability kernel admits a low-rank matrix decomposition \citep{REPUCB22}, and the loss functions may change adversarially but are revealed to the learner at the end of each episode. We propose a policy optimization-based algorithm POLO, and we prove that it attains the $\widetilde{O}(K^{\frac{5}{6}}A^{\frac{1}{2}}d\ln(1+M)/(1-\gamma)^2)$ regret guarantee, where $d$ is rank of the transition kernel (and hence the dimension of the unknown representations), $A$ is the cardinality of the action space, $M$ is the cardinality of the model class, and $\gamma$ is the discounted factor. Notably, our algorithm is oracle-efficient and has a regret guarantee with no dependence on the size of potentially arbitrarily large state space. Furthermore, we also prove an $\Omega(\frac{\gamma^2}{1-\gamma} \sqrt{d A K})$ regret lower bound for this problem, showing that low-rank MDPs are statistically more difficult to learn than linear MDPs in the regret minimization setting. To the best of our knowledge, we present the first algorithm that interleaves representation learning, exploration, and exploitation to achieve the sublinear regret guarantee for RL with nonlinear function approximation and adversarial losses.

In offline reinforcement learning (RL) we have no opportunity to explore so we must make assumptions that the data is sufficient to guide picking a good policy, taking the form of assuming some coverage, realizability, Bellman completeness, and/or hard margin (gap). In this work we propose value-based algorithms for offline RL with PAC guarantees under just partial coverage, specifically, coverage of just a single comparator policy, and realizability of soft (entropy-regularized) Q-function of the single policy and a related function defined as a saddle point of certain minimax optimization problem. This offers refined and generally more lax conditions for offline RL. We further show an analogous result for vanilla Q-functions under a soft margin condition. To attain these guarantees, we leverage novel minimax learning algorithms to accurately estimate soft or vanilla Q-functions with $L^2$-convergence guarantees. Our algorithms' loss functions arise from casting the estimation problems as nonlinear convex optimization problems and Lagrangifying.

ABSTRACT Traffic simulators are used to generate data for learning in intelligent transportation systems (ITSs). A key question is to what extent their modelling assumptions affect the capabilities of ITSs to adapt to various scenarios when deployed in the real world. This work focuses on two simulators commonly used to train reinforcement learning (RL) agents for traffic applications, CityFlow and SUMO. A controlled virtual experiment varying driver behavior and simulation scale finds evidence against distributional equivalence in RL-relevant measures from these simulators, with the root mean squared error and KL divergence being significantly greater than 0 for all assessed measures. While granular real-world validation generally remains infeasible, these findings suggest that traffic simulators are not a deus ex machina for RL training: understanding the impacts of inter-simulator differences is necessary to train and deploy RL-based ITSs. 1 INTRODUCTION Transportation efficiency is becoming an increasingly critical challenge due to continual growth in the volume of people and objects that need to be transported. The 2021 Urban Mobility Report (Schrank et al. 2021) projected that, while the COVID-19 pandemic alleviated congestion, traffic levels in the US will quickly rebound in areas with expanding populations and job markets to produce the most rapid congestion growth since 1982. The increased traffic will stress existing infrastructure and result in social, economic, and environmental costs (Schrank et al. 2021), thus making the development and deployment of intelligent transportation systems (ITSs) a critical priority. At the same time, advances in computational algorithms and roadway infrastructure made in response to these challenges provide opportunities to enhance ITS learning. For example, novel traffic signal control technologies based on reinforcement learning (RL), which learn adaptive signaling policies from simulations generated using real-world traffic data, have already achieved performance on par with and even exceeding traditional control methods (Chen et al. 2020). However, collecting data for ITS learning remains a nontrivial task.