Recent advances in CV and NLP have been largely driven by scaling up the number of network parameters, despite traditional theories suggesting that larger networks are prone to overfitting. These large networks avoid overfitting by integrating components that induce a simplicity bias, guiding models toward simple and generalizable solutions. However, in deep RL, designing and scaling up networks have been less explored. Motivated by this opportunity, we present SimBa, an architecture designed to scale up parameters in deep RL by injecting a simplicity bias. SimBa consists of three components: (i) an observation normalization layer that standardizes inputs with running statistics, (ii) a residual feedforward block to provide a linear pathway from the input to output, and (iii) a layer normalization to control feature magnitudes. By scaling up parameters with SimBa, the sample efficiency of various deep RL algorithms-including off-policy, on-policy, and unsupervised methods-is consistently improved. Moreover, solely by integrating SimBa architecture into SAC, it matches or surpasses state-of-the-art deep RL methods with high computational efficiency across DMC, MyoSuite, and HumanoidBench. These results demonstrate SimBa's broad applicability and effectiveness across diverse RL algorithms and environments.

Meta-reinforcement learning (Meta-RL) has attracted attention due to its capability to enhance reinforcement learning (RL) algorithms, in terms of data efficiency and generalizability. In this paper, we develop a bilevel optimization framework for meta-RL (BO-MRL) to learn the meta-prior for task-specific policy adaptation, which implements multiple-step policy optimization on one-time data collection. Beyond existing meta-RL analyses, we provide upper bounds of the expected optimality gap over the task distribution. This metric measures the distance of the policy adaptation from the learned meta-prior to the task-specific optimum, and quantifies the model's generalizability to the task distribution. We empirically validate the correctness of the derived upper bounds and demonstrate the superior effectiveness of the proposed algorithm over benchmarks.

This study introduces ContextWIN, a novel architecture that extends the Neural Whittle Index Network (NeurWIN) model to address Restless Multi-Armed Bandit (RMAB) problems with a context-aware approach. By integrating a mixture of experts within a reinforcement learning framework, ContextWIN adeptly utilizes contextual information to inform decision-making in dynamic environments, particularly in recommendation systems. A key innovation is the model's ability to assign context-specific weights to a subset of NeurWIN networks, thus enhancing the efficiency and accuracy of the Whittle index computation for each arm. The paper presents a thorough exploration of ContextWIN, from its conceptual foundation to its implementation and potential applications. We delve into the complexities of RMABs and the significance of incorporating context, highlighting how ContextWIN effectively harnesses these elements. The convergence of both the NeurWIN and ContextWIN models is rigorously proven, ensuring theoretical robustness. This work lays the groundwork for future advancements in applying contextual information to complex decision-making scenarios, recognizing the need for comprehensive dataset exploration and environment development for full potential realization.

Effective decision-making in partially observable environments demands robust memory management. Despite their success in supervised learning, current deep-learning memory models struggle in reinforcement learning environments that are partially observable and long-term. They fail to efficiently capture relevant past information, adapt flexibly to changing observations, and maintain stable updates over long episodes. We theoretically analyze the limitations of existing memory models within a unified framework and introduce the Stable Hadamard Memory, a novel memory model for reinforcement learning agents. Our model dynamically adjusts memory by erasing no longer needed experiences and reinforcing crucial ones computationally efficiently. To this end, we leverage the Hadamard product for calibrating and updating memory, specifically designed to enhance memory capacity while mitigating numerical and learning challenges. Our approach significantly outperforms state-of-the-art memory-based methods on challenging partially observable benchmarks, such as meta-reinforcement learning, long-horizon credit assignment, and POPGym, demonstrating superior performance in handling long-term and evolving contexts.

Reinforcement learning (RL) is ubiquitous in the development of modern AI systems. However, state-of-the-art RL agents require extensive, and potentially unsafe, interactions with their environments to learn effectively. These limitations confine RL agents to simulated environments, hindering their ability to learn directly in real-world settings. Despite the notable progress, its application without any use of simulators remains largely limited. This is primarily because, in many cases, RL methods require massive amounts of data for learning while also being inherently unsafe during exploration. In many real-world settings, environments are complex and rarely align exactly with the assumptions made in simulators. Learning directly in the real world allows RL systems to close the sim-to-real gap and continuously adapt to evolving environments and distribution shifts. However, to unlock these advantages, RL algorithms must be sample-efficient and ensure safety throughout the learning process to avoid costly failures or risks in high-stakes applications. For instance, agents learning driving policies in autonomous vehicles must prevent collisions with other cars or pedestrians, even when adapting to new driving environments. This challenge is known as safe exploration, where the agent's exploration is restricted by safety-critical, often unknown, constraints that must be satisfied throughout the learning process . Several works study safe exploration and have demonstrated state-of-the-art performance in terms of both safety and sample efficiency for learning in the real world (Sui et al., 2015; Wischnewski et al., 2019; Berkenkamp et al., 2021; Cooper & Netoff, 2022; Sukhija et al., 2023; Widmer et al., 2023). These methods maintain a "safe set" of policies during learning, selecting policies from this set to safely explore and gradually expand it. Under common regularity assumptions about the constraints, these approaches guarantee safety throughout learning. However, explictily maintaining and expanding a safe set, limits these methods to low-dimensional policies, such as PID controllers. This makes them difficult to scale to more complex tasks such as those considered in deep RL. To this end, we propose a scalable model-based RL algorithm - A CTS AFE - for efficient and safe exploration. Crucially, A CTS AFE learns an uncertainty-aware dynamics model, which it uses to implicitly define and expand the safe set of policies.

The in-context learning (ICL) capability of pre-trained models based on the transformer architecture has received growing interest in recent years. While theoretical understanding has been obtained for ICL in reinforcement learning (RL), the previous results are largely confined to the single-agent setting. This work proposes to further explore the in-context learning capabilities of pre-trained transformer models in competitive multi-agent games, i.e., in-context game-playing (ICGP). Focusing on the classical two-player zero-sum games, theoretical guarantees are provided to demonstrate that pre-trained transformers can provably learn to approximate Nash equilibrium in an in-context manner for both decentralized and centralized learning settings. As a key part of the proof, constructional results are established to demonstrate that the transformer architecture is sufficiently rich to realize celebrated multi-agent game-playing algorithms, in particular, decentralized V-learning and centralized VI-ULCB.

In the real world, unmanned surface vehicles (USV) often need to coordinate with each other to accomplish specific tasks. However, achieving cooperative control in multi-agent systems is challenging due to issues such as non-stationarity and partial observability. Recent advancements in Multi-Agent Reinforcement Learning (MARL) provide new perspectives to address these challenges. Therefore, we propose using the multi-agent deep deterministic policy gradient (MADDPG) algorithm with communication to address multiple ships' cooperation problems under partial observability. We developed two tasks based on OpenAI's gym environment: cooperative navigation and cooperative collision avoidance. In these tasks, ships must not only learn effective control strategies but also establish communication protocols with other agents. We analyze the impact of external noise on communication, the effect of inter-agent communication on performance, and the communication patterns learned by the agents. The results demonstrate that our proposed framework effectively addresses cooperative navigation and collision avoidance among multiple vessels, significantly outperforming traditional single-agent algorithms. Agents establish a consistent communication protocol, enabling them to compensate for missing information through shared observations and achieve better coordination.

Within batch reinforcement learning, safe policy improvement (SPI) seeks to ensure that the learnt policy performs at least as well as the behavior policy that generated the dataset. The core challenge in SPI is seeking improvements while balancing risk when many state-action pairs may be infrequently visited. In this work, we introduce Decision Points RL (DPRL), an algorithm that restricts the set of state-action pairs (or regions for continuous states) considered for improvement. DPRL ensures high-confidence improvement in densely visited states (i.e. decision points) while still utilizing data from sparsely visited states. By appropriately limiting where and how we may deviate from the behavior policy, we achieve tighter bounds than prior work; specifically, our data-dependent bounds do not scale with the size of the state and action spaces. In addition to the analysis, we demonstrate that DPRL is both safe and performant on synthetic and real datasets.

With the explosive growth of machine learning techniques and applications, new paradigms and models with transformative power are enriching the field. One of the most remarkable trends in recent years is the rapid rise of significance of Riemannian geometry and Lie group theory. The underlying cause is the rising complexity of the data, motivating more sophisticated approaches, thus leading to the wide recognition that a great deal of data sets exhibit an intrinsic curvature. In other words, many data sets are naturally represented or faithfully embedded into non-Euclidean spaces. One apparent example of this kind are rotational motions in robotics.

This work introduces a novel off-policy Reinforcement Learning (RL) algorithm that utilizes a transformer-based architecture for predicting the state-action values for a sequence of actions. These returns are effectively used to update the policy that predicts a smooth trajectory instead of a single action in each decision step. Predicting a whole trajectory of actions is commonly done in episodic RL (ERL) (Kober & Peters, 2008) and differs conceptually from conventional step-based RL (SRL) methods like PPO (Schulman et al., 2017) and SAC (Haarnoja et al., 2018a) where an action is sampled in each time step. The action selection concept in ERL is promising as shown in recent works in RL (Otto et al., 2022; Li et al., 2024). Similar insights have been made in the field of Imitation Learning, where predicting action sequences instead of single actions has led to great success (Zhao et al., 2023; Reuss et al., 2024). Additionally, decision-making in ERL aligns with the human's decision-making strategy, where the human generally does not decide in each single time step but rather performs a whole sequence of actions to complete a task - for instance, swinging an arm to play tennis without overthinking each per-step movement. Episodic RL is a distinct family of RL that emphasizes the maximization of returns over entire episodes, typically lasting several seconds, rather than optimizing the intermediate states during environment interactions (Whitley et al., 1993; Igel, 2003; Peters & Schaal, 2008). Unlike SRL, ERL shifts the solution search from per-step actions to a parameterized trajectory space, leveraging techniques like Movement Primitives (MPs) (Schaal, 2006; Paraschos et al., 2013) for generating action sequences. This approach enables a broader exploration horizon (Kober & Peters, 2008), captures temporal and degrees of freedom (DoF) correlations (Li et al., 2024), and ensures smooth transitions between re-planning phases (Otto et al., 2023).