Maximum entropy (MaxEnt) RL maximizes a combination of the original task reward and an entropy reward. It is believed that the regularization imposed by entropy, on both policy improvement and policy evaluation, together contributes to good exploration, training convergence, and robustness of learned policies. This paper takes a closer look at entropy as an intrinsic reward, by conducting various ablation studies on soft actor-critic (SAC), a popular representative of MaxEnt RL. Our findings reveal that in general, entropy rewards should be applied with caution to policy evaluation. On one hand, the entropy reward, like any other intrinsic reward, could obscure the main task reward if it is not properly managed. We identify some failure cases of the entropy reward especially in episodic Markov decision processes (MDPs), where it could cause the policy to be overly optimistic or pessimistic. On the other hand, our large-scale empirical study shows that using entropy regularization alone in policy improvement, leads to comparable or even better performance and robustness than using it in both policy improvement and policy evaluation. Based on these observations, we recommend either normalizing the entropy reward to a zero mean (SACZero), or simply removing it from policy evaluation (SACLite) for better practical results.

We consider the safe reinforcement learning (RL) problem of maximizing utility while satisfying provided constraints. Since we do not assume any prior knowledge or pre-training of the safety concept, we are interested in asymptotic constraint satisfaction. A popular approach in this line of research is to combine the Lagrangian method with a model-free RL algorithm to adjust the weight of the constraint reward dynamically. It relies on a single policy to handle the conflict between utility and constraint rewards, which is often challenging. Inspired by the safety layer design (Dalal et al., 2018), we propose to separately learn a safety editor policy that transforms potentially unsafe actions output by a utility maximizer policy into safe ones. The safety editor is trained to maximize the constraint reward while minimizing a hinge loss of the utility Q values of actions before and after the edit. On 12 custom Safety Gym (Ray et al., 2019) tasks and 2 safe racing tasks with very harsh constraint thresholds, our approach demonstrates outstanding utility performance while complying with the constraints. Ablation studies reveal that our two-policy design is critical. Simply doubling the model capacity of typical single-policy approaches will not lead to comparable results. The Q hinge loss is also important in certain circumstances, and replacing it with the usual L2 distance could fail badly.

The classical Policy Iteration (PI) algorithm alternates between greedy one-step policy improvement and policy evaluation. Recent literature shows that multi-step lookahead policy improvement leads to a better convergence rate at the expense of increased complexity per iteration. However, prior to running the algorithm, one cannot tell what is the best fixed lookahead horizon. Moreover, per a given run, using a lookahead of horizon larger than one is often wasteful. In this work, we propose for the first time to dynamically adapt the multi-step lookahead horizon as a function of the state and of the value estimate. We devise two PI variants and analyze the trade-off between iteration count and computational complexity per iteration. The first variant takes the desired contraction factor as the objective and minimizes the per-iteration complexity. The second variant takes as input the computational complexity per iteration and minimizes the overall contraction factor. We then devise a corresponding DQN-based algorithm with an adaptive tree search horizon. We also include a novel enhancement for on-policy learning: per-depth value function estimator. Lastly, we demonstrate the efficacy of our adaptive lookahead method in a maze environment and in Atari.

Humans process visual information with varying resolution (foveated visual system) and explore images by orienting through eye movements the high-resolution fovea to points of interest. The Bayesian ideal searcher (IS) that employs complete knowledge of task-relevant information optimizes eye movement strategy and achieves the optimal search performance. The IS can be employed as an important tool to evaluate the optimality of human eye movements, and potentially provide guidance to improve human observer visual search strategies. Najemnik and Geisler (2005) derived an IS for backgrounds of spatial 1/f noise. The corresponding template responses follow Gaussian distributions and the optimal search strategy can be analytically determined. However, the computation of the IS can be intractable when considering more realistic and complex backgrounds such as medical images. Modern reinforcement learning methods, successfully applied to obtain optimal policy for a variety of tasks, do not require complete knowledge of the background generating functions and can be potentially applied to anatomical backgrounds. An important first step is to validate the optimality of the reinforcement learning method. In this study, we investigate the ability of a reinforcement learning method that employs Q-network to approximate the IS. We demonstrate that the search strategy corresponding to the Q-network is consistent with the IS search strategy. The findings show the potential of the reinforcement learning with Q-network approach to estimate optimal eye movement planning with real anatomical backgrounds.

Safe Policy Improvement (SPI) aims at provable guarantees that a learned policy is at least approximately as good as a given baseline policy. Building on SPI with Soft Baseline Bootstrapping (Soft-SPIBB) by Nadjahi et al., we identify theoretical issues in their approach, provide a corrected theory, and derive a new algorithm that is provably safe on finite Markov Decision Processes (MDP). Additionally, we provide a heuristic algorithm that exhibits the best performance among many state of the art SPI algorithms on two different benchmarks. Furthermore, we introduce a taxonomy of SPI algorithms and empirically show an interesting property of two classes of SPI algorithms: while the mean performance of algorithms that incorporate the uncertainty as a penalty on the action-value is higher, actively restricting the set of policies more consistently produces good policies and is, thus, safer.

Enabling reinforcement learning (RL) agents to leverage a knowledge base while learning from experience promises to advance RL in knowledge intensive domains. However, it has proven difficult to leverage knowledge that is not manually tailored to the environment. We propose to use the subclass relationships present in open-source knowledge graphs to abstract away from specific objects. We develop a residual policy gradient method that is able to integrate knowledge across different abstraction levels in the class hierarchy. Our method results in improved sample efficiency and generalisation to unseen objects in commonsense games, but we also investigate failure modes, such as excessive noise in the extracted class knowledge or environments with little class structure.

Fine-tuning reinforcement learning (RL) models has been challenging because of a lack of large scale off-the-shelf datasets as well as high variance in transferability among different environments. Recent work has looked at tackling offline RL from the perspective of sequence modeling with improved results as result of the introduction of the Transformer architecture. However, when the model is trained from scratch, it suffers from slow convergence speeds. In this paper, we look to take advantage of this formulation of reinforcement learning as sequence modeling and investigate the transferability of pre-trained sequence models on other domains (vision, language) when finetuned on offline RL tasks (control, games). To this end, we also propose techniques to improve transfer between these domains. Results show consistent performance gains in terms of both convergence speed and reward on a variety of environments, accelerating training by 3-6x and achieving state-of-the-art performance in a variety of tasks using Wikipedia-pretrained and GPT2 language models. We hope that this work not only brings light to the potentials of leveraging generic sequence modeling techniques and pre-trained models for RL, but also inspires future work on sharing knowledge between generative modeling tasks of completely different domains.

Hierarchical Imitation Learning (HIL) is an effective way for robots to learn sub-skills from long-horizon unsegmented demonstrations. However, the learned hierarchical structure lacks the mechanism to transfer across multi-tasks or to new tasks, which makes them have to learn from scratch when facing a new situation. Transferring and reorganizing modular sub-skills require fast adaptation ability of the whole hierarchical structure. In this work, we propose Dual Meta Imitation Learning (DMIL), a hierarchical meta imitation learning method where the high-level network and sub-skills are iteratively meta-learned with model-agnostic meta-learning. DMIL uses the likelihood of state-action pairs from each sub-skill as the supervision for the high-level network adaptation, and use the adapted high-level network to determine different data set for each sub-skill adaptation. We theoretically prove the convergence of the iterative training process of DMIL and establish the connection between DMIL and Expectation-Maximization algorithm. Empirically, we achieve state-of-the-art few-shot imitation learning performance on the Meta-world \cite{metaworld} benchmark and competitive results on long-horizon tasks of Kitchen environments.

Planning based on long and short term time series forecasts is a common practice across many industries. In this context, temporal aggregation and reconciliation techniques have been useful in improving forecasts, reducing model uncertainty, and providing a coherent forecast across different time horizons. However, an underlying assumption spanning all these techniques is the complete availability of data across all levels of the temporal hierarchy, while this offers mathematical convenience but most of the time low frequency data is partially completed and it is not available while forecasting. On the other hand, high frequency data can significantly change in a scenario like the COVID pandemic and this change can be used to improve forecasts that will otherwise significantly diverge from long term actuals. We propose a dynamic reconciliation method whereby we formulate the problem of informing low frequency forecasts based on high frequency actuals as a Markov Decision Process (MDP) allowing for the fact that we do not have complete information about the dynamics of the process. This allows us to have the best long term estimates based on the most recent data available even if the low frequency cycles have only been partially completed. The MDP has been solved using a Time Differenced Reinforcement learning (TDRL) approach with customizable actions and improves the long terms forecasts dramatically as compared to relying solely on historical low frequency data. The result also underscores the fact that while low frequency forecasts can improve the high frequency forecasts as mentioned in the temporal reconciliation literature (based on the assumption that low frequency forecasts have lower noise to signal ratio) the high frequency forecasts can also be used to inform the low frequency forecasts.

In this work, we study the use of the Bellman equation as a surrogate objective for value prediction accuracy. While the Bellman equation is uniquely solved by the true value function over all state-action pairs, we find that the Bellman error (the difference between both sides of the equation) is a poor proxy for the accuracy of the value function. In particular, we show that (1) due to cancellations from both sides of the Bellman equation, the magnitude of the Bellman error is only weakly related to the distance to the true value function, even when considering all state-action pairs, and (2) in the finite data regime, the Bellman equation can be satisfied exactly by infinitely many suboptimal solutions. This means that the Bellman error can be minimized without improving the accuracy of the value function. We demonstrate these phenomena through a series of propositions, illustrative toy examples, and empirical analysis in standard benchmark domains.