We consider the reinforcement learning problem for partially observed Markov decision processes (POMDPs) with large or even countably infinite state spaces, where the controller has access to only noisy observations of the underlying controlled Markov chain. We consider a natural actor-critic method that employs a finite internal memory for policy parameterization, and a multi-step temporal difference learning algorithm for policy evaluation. We establish, to the best of our knowledge, the first non-asymptotic global convergence of actor-critic methods for partially observed systems under function approximation. In particular, in addition to the function approximation and statistical errors that also arise in MDPs, we explicitly characterize the error due to the use of finite-state controllers. This additional error is stated in terms of the total variation distance between the traditional belief state in POMDPs and the posterior distribution of the hidden state when using a finite-state controller. Further, we show that this error can be made small in the case of sliding-block controllers by using larger block sizes.

The use of machine learning in algorithmic trading systems is increasingly common. In a typical set-up, supervised learning is used to predict the future prices of assets, and those predictions drive a simple trading and execution strategy. This is quite effective when the predictions have sufficient signal, markets are liquid, and transaction costs are low. However, those conditions often do not hold in thinly traded financial markets and markets for differentiated assets such as real estate or vehicles. In these markets, the trading strategy must consider the long-term effects of taking positions that are relatively more difficult to change. In this work, we propose a Reinforcement Learning (RL) algorithm that trades based on signals from a learned predictive model and addresses these challenges. We test our algorithm on 20+ years of equity data from Bursa Malaysia.

In recent years, a new architecture for vision-based tasks that does not use convolutions called the Vision Transformer (ViT) (Dosovitskiy et al., 2020) has shown impressive results in several benchmarks. This architecture presents much weaker inductive biases when compared to a CNN, which can result in lower data efficiency. The Vision Transformer, unlike the CNNs, can capture relations between parts of an image (patches) that are far apart from each other, thus deriving global information that can help the model perform better in certain tasks. When the model is pretrained, using supervised or self-supervised learning, it manages to surpass in some cases the best convolution-based models in terms of task performance. Nonetheless, despite the successes obtained in computer vision these results are yet to be seen in reinforcement learning. Moreover, while some areas of machine learning have transitioned to large pretrained models, current Deep RL research is still largely based on small neural networks that are trained from tabula rasa. Despite the successes of deep reinforcement learning agents in the last decade, these still require a large amount of data or interactions to learn good policies. This data inefficiency makes current methods difficult to apply to environments where interactions are more expensive or data is scarce, which is the case in many real-world applications. In environments where the agent does not have full access to the current state, i.e. partially observable environments, this problem becomes even more prominent, since the agent not only needs to learn the state-to-action mapping but also a state representation function that tries to be informative about the current state given an observation.

Mitchell Barrington Center for AI Safety University of Michigan University of Southern California Abstract This paper argues that training AI systems with absolute constraints--which forbid certain acts irrespective of the amount of value they might produce--may make considerable progress on many AI safety problems in principle. First, it provides a guardrail for avoiding the very worst outcomes of misalignment: An AI attempting to commit mass murder might have correctly deduced that doing so maximizes expected value, but more likely, the system is severely misaligned. Second, it could prevent AIs from causing catastrophes for the sake of very valuable consequences, such as replacing humans with a much larger number of beings living at a higher welfare level. Third, it makes systems more corrigible, allowing creators to make corrective interventions in them, such as altering their objective functions or shutting them down. And fourth, it helps systems explore their environment more safely by prohibiting them from exploring especially dangerous acts. I offer a decision-theoretic formalization of an absolute constraints, improving on existing models in the literature, and use this model to prove some results about the training and behavior of absolutist AIs. I conclude by showing that, although absolutist AIs will not maximize expected value, they will not be susceptible to behave irrationally, and they will not (contra coherence arguments) face environmental pressure to become expected-value maximizers. Introduction Advanced AI systems are expected to be dangerous because of the opacity of their goals: We may know that they will effectively pursue their goals but fail to know what those goals are.

Preference-based reinforcement learning (PbRL) promises to learn a complex reward function with binary human preference. However, such human-in-the-loop formulation requires considerable human effort to assign preference labels to segment pairs, hindering its large-scale applications. Recent approache has tried to reuse unlabeled segments, which implicitly elucidates the distribution of segments and thereby alleviates the human effort. And consistency regularization is further considered to improve the performance of semi-supervised learning. However, we notice that, unlike general classification tasks, in PbRL there exits a unique phenomenon that we defined as similarity trap in this paper. Intuitively, human can have diametrically opposite preferredness for similar segment pairs, but such similarity may trap consistency regularization fail in PbRL. Due to the existence of similarity trap, such consistency regularization improperly enhances the consistency possiblity of the model's predictions between segment pairs, and thus reduces the confidence in reward learning, since the augmented distribution does not match with the original one in PbRL. To overcome such issue, we present a self-training method along with our proposed peer regularization, which penalizes the reward model memorizing uninformative labels and acquires confident predictions. Empirically, we demonstrate that our approach is capable of learning well a variety of locomotion and robotic manipulation behaviors using different semi-supervised alternatives and peer regularization.

Language Models and Vision Language Models have recently demonstrated unprecedented capabilities in terms of understanding human intentions, reasoning, scene understanding, and planning-like behaviour, in text form, among many others. In this work, we investigate how to embed and leverage such abilities in Reinforcement Learning (RL) agents. We design a framework that uses language as the core reasoning tool, exploring how this enables an agent to tackle a series of fundamental RL challenges, such as efficient exploration, reusing experience data, scheduling skills, and learning from observations, which traditionally require separate, vertically designed algorithms. We test our method on a sparse-reward simulated robotic manipulation environment, where a robot needs to stack a set of objects. We demonstrate substantial performance improvements over baselines in exploration efficiency and ability to reuse data from offline datasets, and illustrate how to reuse learned skills to solve novel tasks or imitate videos of human experts. In recent years, the literature has seen a series of remarkable Deep Learning (DL) success stories (3), with breakthroughs particularly in the fields of Natural Language Processing (4; 19; 8; 29) and Computer Vision (2; 25; 36; 37).

The ability to model the mental states of others is crucial to human social intelligence, and can offer similar benefits to artificial agents with respect to the social dynamics induced in multi-agent settings. We present a method of grounding semantically meaningful, human-interpretable beliefs within policies modeled by deep networks. We then consider the task of 2nd-order belief prediction. We propose that ability of each agent to predict the beliefs of the other agents can be used as an intrinsic reward signal for multi-agent reinforcement learning. Finally, we present preliminary empirical results in a mixed cooperative-competitive environment.

Maximizing long-term rewards is the primary goal in sequential decision-making problems. The majority of existing methods assume that side information is freely available, enabling the learning agent to observe all features' states before making a decision. In real-world problems, however, collecting beneficial information is often costly. That implies that, besides individual arms' reward, learning the observations of the features' states is essential to improve the decision-making strategy. The problem is aggravated in a non-stationary environment where reward and cost distributions undergo abrupt changes over time. To address the aforementioned dual learning problem, we extend the contextual bandit setting and allow the agent to observe subsets of features' states. The objective is to maximize the long-term average gain, which is the difference between the accumulated rewards and the paid costs on average. Therefore, the agent faces a trade-off between minimizing the cost of information acquisition and possibly improving the decision-making process using the obtained information. To this end, we develop an algorithm that guarantees a sublinear regret in time. Numerical results demonstrate the superiority of our proposed policy in a real-world scenario.

In many reinforcement learning tasks, the agent has to learn to interact with many objects of different types and generalize to unseen combinations and numbers of objects. Often a task is a composition of previously learned tasks (e.g. block stacking). These are examples of compositional generalization, in which we compose object-centric representations to solve complex tasks. Recent works have shown the benefits of object-factored representations and hierarchical abstractions for improving sample efficiency in these settings. On the other hand, these methods do not fully exploit the benefits of factorization in terms of object attributes. In this paper, we address this opportunity and introduce the Dynamic Attribute FacTored RL (DAFT-RL) framework. In DAFT-RL, we leverage object-centric representation learning to extract objects from visual inputs. We learn to classify them in classes and infer their latent parameters. For each class of object, we learn a class template graph that describes how the dynamics and reward of an object of this class factorize according to its attributes. We also learn an interaction pattern graph that describes how objects of different classes interact with each other at the attribute level. Through these graphs and a dynamic interaction graph that models the interactions between objects, we can learn a policy that can then be directly applied in a new environment by just estimating the interactions and latent parameters. We evaluate DAFT-RL in three benchmark datasets and show our framework outperforms the state-of-the-art in generalizing across unseen objects with varying attributes and latent parameters, as well as in the composition of previously learned tasks.

Boolean satisfiability (SAT) is a fundamental NP-complete problem with many applications, including automated planning and scheduling. To solve large instances, SAT solvers have to rely on heuristics, e.g., choosing a branching variable in DPLL and CDCL solvers. Such heuristics can be improved with machine learning (ML) models; they can reduce the number of steps but usually hinder the running time because useful models are relatively large and slow. We suggest the strategy of making a few initial steps with a trained ML model and then releasing control to classical heuristics; this simplifies cold start for SAT solving and can decrease both the number of steps and overall runtime, but requires a separate decision of when to release control to the solver. Moreover, we introduce a modification of Graph-Q-SAT tailored to SAT problems converted from other domains, e.g., open shop scheduling problems. We validate the feasibility of our approach with random and industrial SAT problems.