Learning Task-relevant Sequence Representations via Intrinsic Dynamics Characteristics in Reinforcement Learning

Learning task-relevant state representations is crucial to solving the problem of scene generalization in visual deep reinforcement learning. Prior work typically establishes a self-supervised auxiliary learner, introducing elements (e.g., rewards and actions) to extract task-relevant state information from observations through behavioral similarity metrics. However, the methods often ignore the inherent relationships between the elements (e.g., dynamics relationships) that are essential for learning accurate representations, and they are also limited to single-step metrics, which impedes the discrimination of short-term similar task/behavior information in long-term dynamics transitions. To solve the issues, we propose an intrinsic dynamic characteristics-driven sequence representation learning method (DSR) over a common DRL frame. Concretely, inspired by the fact of state transition in the underlying system, it constrains the optimization of the encoder via modeling the dynamics equations related to the state transition, which prompts the latent encoding information to satisfy the state transition process and thereby distinguishes state space and noise space. Further, to refine the ability of encoding similar tasks based on dynamics constraints, DSR also sequentially models inherent dynamics equation relationships from the perspective of sequence elements' frequency domain and multi-step prediction. Finally, experimental results show that DSR has achieved a significant performance boost in the Distracting DMControl Benchmark, with an average of 78.9% over the backbone baseline. Further results indicate that it also achieves the best performance in real-world autonomous driving tasks in the CARLA simulator. Moreover, the qualitative analysis results of t-SNE visualization validate that our method possesses superior representation ability on visual tasks.

Episodic Reinforcement Learning with Expanded State-reward Space

Empowered by deep neural networks, deep reinforcement learning (DRL) has demonstrated tremendous empirical successes in various domains, including games, health care, and autonomous driving. Despite these advancements, DRL is still identified as data-inefficient as effective policies demand vast numbers of environmental samples. Recently, episodic control (EC)-based model-free DRL methods enable sample efficiency by recalling past experiences from episodic memory. However, existing EC-based methods suffer from the limitation of potential misalignment between the state and reward spaces for neglecting the utilization of (past) retrieval states with extensive information, which probably causes inaccurate value estimation and degraded policy performance. To tackle this issue, we introduce an efficient EC-based DRL framework with expanded state-reward space, where the expanded states used as the input and the expanded rewards used in the training both contain historical and current information. To be specific, we reuse the historical states retrieved by EC as part of the input states and integrate the retrieved MC-returns into the immediate reward in each interactive transition. As a result, our method is able to simultaneously achieve the full utilization of retrieval information and the better evaluation of state values by a Temporal Difference (TD) loss. Empirical results on challenging Box2d and Mujoco tasks demonstrate the superiority of our method over a recent sibling method and common baselines. Further, we also verify our method's effectiveness in alleviating Q-value overestimation by additional experiments of Q-value comparison.

Sequential Action-Induced Invariant Representation for Reinforcement Learning

How to accurately learn task-relevant state representations from high-dimensional observations with visual distractions is a realistic and challenging problem in visual reinforcement learning. Recently, unsupervised representation learning methods based on bisimulation metrics, contrast, prediction, and reconstruction have shown the ability for task-relevant information extraction. However, due to the lack of appropriate mechanisms for the extraction of task information in the prediction, contrast, and reconstruction-related approaches and the limitations of bisimulation-related methods in domains with sparse rewards, it is still difficult for these methods to be effectively extended to environments with distractions. To alleviate these problems, in the paper, the action sequences, which contain task-intensive signals, are incorporated into representation learning. Specifically, we propose a Sequential Action--induced invariant Representation (SAR) method, in which the encoder is optimized by an auxiliary learner to only preserve the components that follow the control signals of sequential actions, so the agent can be induced to learn the robust representation against distractions. We conduct extensive experiments on the DeepMind Control suite tasks with distractions while achieving the best performance over strong baselines. We also demonstrate the effectiveness of our method at disregarding task-irrelevant information by deploying SAR to real-world CARLA-based autonomous driving with natural distractions. Finally, we provide the analysis results of generalization drawn from the generalization decay and t-SNE visualization. Code and demo videos are available at https://github.com/DMU-XMU/SAR.git.

Online Learning and Planning in Partially Observable Domains without Prior Knowledge

How an agent can act optimally in stochastic, partially observable domains is a challenge problem, the standard approach to address this issue is to learn the domain model firstly and then based on the learned model to find the (near) optimal policy. However, offline learning the model often needs to store the entire training data and cannot utilize the data generated in the planning phase. Furthermore, current research usually assumes the learned model is accurate or presupposes knowledge of the nature of the unobservable part of the world. In this paper, for systems with discrete settings, with the benefits of Predictive State Representations~(PSRs), a model-based planning approach is proposed where the learning and planning phases can both be executed online and no prior knowledge of the underlying system is required. Experimental results show compared to the state-of-the-art approaches, our algorithm achieved a high level of performance with no prior knowledge provided, along with theoretical advantages of PSRs. Source code is available at https://github.com/DMU-XMU/PSR-MCTS-Online.

Combining Offline Models and Online Monte-Carlo Tree Search for Planning from Scratch

Planning in stochastic and partially observable environments is a central issue in artificial intelligence. One commonly used technique for solving such a problem is by constructing an accurate model firstly. Although some recent approaches have been proposed for learning optimal behaviour under model uncertainty, prior knowledge about the environment is still needed to guarantee the performance of the proposed algorithms. With the benefits of the Predictive State Representations~(PSRs) approach for state representation and model prediction, in this paper, we introduce an approach for planning from scratch, where an offline PSR model is firstly learned and then combined with online Monte-Carlo tree search for planning with model uncertainty. By comparing with the state-of-the-art approach of planning with model uncertainty, we demonstrated the effectiveness of the proposed approaches along with the proof of their convergence. The effectiveness and scalability of our proposed approach are also tested on the RockSample problem, which are infeasible for the state-of-the-art BA-POMDP based approaches.

Conformation Clustering of Long MD Protein Dynamics with an Adversarial Autoencoder

Recent developments in specialized computer hardware have greatly accelerated atomic level Molecular Dynamics (MD) simulations. A single GPU-attached cluster is capable of producing microsecond-length trajectories in reasonable amounts of time. Multiple protein states and a large number of microstates associated with folding and with the function of the protein can be observed as conformations sampled in the trajectories. Clustering those conformations, however, is needed for identifying protein states, evaluating transition rates and understanding protein behavior. In this paper, we propose a novel data-driven generative conformation clustering method based on the adversarial autoencoder (AAE) and provide the associated software implementation Cong. The method was tested using a 208 microseconds MD simulation of the fast-folding peptide Trp-Cage (20 residues) obtained from the D.E. Shaw Research Group. The proposed clustering algorithm identifies many of the salient features of the folding process by grouping a large number of conformations that share common features not easily identifiable in the trajectory.

Selecting Bases in Spectral learning of Predictive State Representations via Model Entropy

Predictive State Representations (PSRs) are powerful techniques for modelling dynamical systems, which represent a state as a vector of predictions about future observable events (tests). In PSRs, one of the fundamental problems is the learning of the PSR model of the underlying system. Recently, spectral methods have been successfully used to address this issue by treating the learning problem as the task of computing an singular value decomposition (SVD) over a submatrix of a special type of matrix called the Hankel matrix. Under the assumptions that the rows and columns of the submatrix of the Hankel Matrix are sufficient~(which usually means a very large number of rows and columns, and almost fails in practice) and the entries of the matrix can be estimated accurately, it has been proven that the spectral approach for learning PSRs is statistically consistent and the learned parameters can converge to the true parameters. However, in practice, due to the limit of the computation ability, only a finite set of rows or columns can be chosen to be used for the spectral learning. While different sets of columns usually lead to variant accuracy of the learned model, in this paper, we propose an approach for selecting the set of columns, namely basis selection, by adopting a concept of model entropy to measure the accuracy of the learned model. Experimental results are shown to demonstrate the effectiveness of the proposed approach.