Transfer learning in Reinforcement Learning (RL) enables agents to leverage knowledge from source tasks to accelerate learning in target tasks. While prior work, such as the Attend, Adapt, and Transfer (A2T) framework, addresses negative transfer and selective transfer, other critical challenges remain underexplored. This paper introduces the Generalized Adaptive Transfer Network (GATN), a deep RL architecture designed to tackle task generalization across domains, robustness to environmental changes, and computational efficiency in transfer. GATN employs a domain-agnostic representation module, a robustness-aware policy adapter, and an efficient transfer scheduler to achieve these goals. We evaluate GATN on diverse benchmarks, including Atari 2600, MuJoCo, and a custom chatbot dialogue environment, demonstrating superior performance in cross-domain generalization, resilience to dynamic environments, and reduced computational overhead compared to baselines. Our findings suggest GATN is a versatile framework for real-world RL applications, such as adaptive chatbots and robotic control.
Deep Reinforcement Learning (DRL) has achieved remarkable success in complex sequential decision-making tasks, such as playing Atari 2600 games and mastering board games. A critical yet underexplored aspect of DRL is the temporal scale of action execution. We propose a novel paradigm that integrates contextual bandits with DRL to adaptively select action durations, enhancing policy flexibility and computational efficiency. Our approach augments a Deep Q-Network (DQN) with a contextual bandit module that learns to choose optimal action repetition rates based on state contexts. Experiments on Atari 2600 games demonstrate significant performance improvements over static duration baselines, highlighting the efficacy of adaptive temporal abstractions in DRL. This paradigm offers a scalable solution for real-time applications like gaming and robotics, where dynamic action durations are critical.
One proposed mechanism to improve exploration in reinforcement learning is through the use of macro-actions. Paradoxically though, in many scenarios the naive addition of macro-actions does not lead to better exploration, but rather the opposite. It has been argued that this was caused by adding non-useful macros and multiple works have focused on mechanisms to discover effectively environment-specific useful macros. In this work, we take a slightly different perspective. We argue that the difficulty stems from the trade-offs between reducing the average number of decisions per episode versus increasing the size of the action space. Namely, one typically treats each potential macro-action as independent and atomic, hence strictly increasing the search space and making typical exploration strategies inefficient. To address this problem we propose a novel regularization term that exploits the relationship between actions and macro-actions to improve the credit assignment mechanism by reducing the effective dimension of the action space and, therefore, improving exploration. The term relies on a similarity matrix that is meta-learned jointly with learning the desired policy. We empirically validate our strategy looking at macro-actions in Atari games, and the StreetFighter II environment. Our results show significant improvements over the Rainbow-DQN baseline in all environments. Additionally, we show that the macro-action similarity is transferable to related environments. We believe this work is a small but important step towards understanding how the similarity-imposed geometry on the action space can be exploited to improve credit assignment and exploration, therefore making learning more effective.
We present TextAtari, a benchmark for evaluating language agents on very long-horizon decision-making tasks spanning up to 100,000 steps. By translating the visual state representations of classic Atari games into rich textual descriptions, TextAtari creates a challenging test bed that bridges sequential decision-making with natural language processing. The benchmark includes nearly 100 distinct tasks with varying complexity, action spaces, and planning horizons, all rendered as text through an unsupervised representation learning framework (AtariARI). We evaluate three open-source large language models (Qwen2.5-7B, Gemma-7B, and Llama3.1-8B) across three agent frameworks (zero-shot, few-shot chain-of-thought, and reflection reasoning) to assess how different forms of prior knowledge affect performance on these long-horizon challenges. Four scenarios-Basic, Obscured, Manual Augmentation, and Reference-based-investigate the impact of semantic understanding, instruction comprehension, and expert demonstrations on agent decision-making. Our results reveal significant performance gaps between language agents and human players in extensive planning tasks, highlighting challenges in sequential reasoning, state tracking, and strategic planning across tens of thousands of steps. TextAtari provides standardized evaluation protocols, baseline implementations, and a framework for advancing research at the intersection of language models and planning.
One of the primary aspirations in reinforcement learning research is developing general-purpose agents capable of rapidly adapting to and mastering novel tasks. While RL gaming agents have mastered many Atari games, they remain slow and costly to train for each game. In this work, we demonstrate that latest reasoning LLMs with out-of-domain RL post-training can play a challenging Atari game called Frogger under a zero-shot setting. We then investigate the effect of in-context learning and the amount of reasoning effort on LLM performance. Lastly, we demonstrate a way to bootstrap traditional RL method with LLM demonstrations, which significantly improves their performance and sample efficiency. Our implementation is open sourced at https://github.com/AlienKevin/frogger.
Recently, deep reinforcement learning has emerged as a promising approach for solving complex combinatorial optimization problems. Broadly, deep reinforcement learning methods fall into two categories: policy-based and value-based. While value-based approaches have achieved notable success in domains such as the Arcade Learning Environment, the combinatorial optimization community has predominantly favored policy-based methods, often overlooking the potential of value-based algorithms. In this work, we conduct a comprehensive empirical evaluation of value-based algorithms, including the deep q-network and several of its advanced extensions, within the context of two complex combinatorial problems: the job-shop and the flexible job-shop scheduling problems, two fundamental challenges with multiple industrial applications. Our results challenge the assumption that policy-based methods are inherently superior for combinatorial optimization. We show that several value-based approaches can match or even outperform the widely adopted proximal policy optimization algorithm, suggesting that value-based strategies deserve greater attention from the combinatorial optimization community. Our code is openly available at: https://github.com/AJ-Correa/Unraveling-the-Rainbow.
Current approaches in Explainable Deep Reinforcement Learning have limitations in which the attention mask has a displacement with the objects in visual input. This work addresses a spatial problem within traditional Convolutional Neural Networks (CNNs). We propose the Interpretable Feature Extractor (IFE) architecture, aimed at generating an accurate attention mask to illustrate both "what" and "where" the agent concentrates on in the spatial domain. Our design incorporates a Human-Understandable Encoding module to generate a fully interpretable attention mask, followed by an Agent-Friendly Encoding module to enhance the agent's learning efficiency. These two components together form the Interpretable Feature Extractor for vision-based deep reinforcement learning to enable the model's interpretability. The resulting attention mask is consistent, highly understandable by humans, accurate in spatial dimension, and effectively highlights important objects or locations in visual input. The Interpretable Feature Extractor is integrated into the Fast and Data-efficient Rainbow framework, and evaluated on 57 ATARI games to show the effectiveness of the proposed approach on Spatial Preservation, Interpretability, and Data-efficiency. Finally, we showcase the versatility of our approach by incorporating the IFE into the Asynchronous Advantage Actor-Critic Model.
Planning with options -- a sequence of primitive actions -- has been shown effective in reinforcement learning within complex environments. Previous studies have focused on planning with predefined options or learned options through expert demonstration data. Inspired by MuZero, which learns superhuman heuristics without any human knowledge, we propose a novel approach, named OptionZero. OptionZero incorporates an option network into MuZero, providing autonomous discovery of options through self-play games. Furthermore, we modify the dynamics network to provide environment transitions when using options, allowing searching deeper under the same simulation constraints. Empirical experiments conducted in 26 Atari games demonstrate that OptionZero outperforms MuZero, achieving a 131.58% improvement in mean human-normalized score. Our behavior analysis shows that OptionZero not only learns options but also acquires strategic skills tailored to different game characteristics. Our findings show promising directions for discovering and using options in planning. Our code is available at https://rlg.iis.sinica.edu.tw/papers/optionzero.
World model (WM) agents enable sample-efficient reinforcement learning by learning policies entirely from simulated experience. However, existing token-based world models (TBWMs) are limited to visual inputs and discrete actions, restricting their adoption and applicability. Moreover, although both intrinsic motivation and prioritized WM replay have shown promise in improving WM performance and generalization, they remain underexplored in this setting, particularly in combination. We introduce Simulus, a highly modular TBWM agent that integrates (1) a modular multi-modality tokenization framework, (2) intrinsic motivation, (3) prioritized WM replay, and (4) regression-as-classification for reward and return prediction. Simulus achieves state-of-the-art sample efficiency for planning-free WMs across three diverse benchmarks. Ablation studies reveal the individual contribution of each component while highlighting their synergy. Our code and model weights are publicly available at https://github.com/leor-c/Simulus.
In deep reinforcement learning, policy optimization methods need to deal with issues such as function approximation and the reuse of off-policy data. Standard policy gradient methods do not handle off-policy data well, leading to premature convergence and instability. This paper introduces a method to stabilize policy optimization when off-policy data are reused. The idea is to include a Bregman divergence between the behavior policy that generates the data and the current policy to ensure small and safe policy updates with off-policy data. The Bregman divergence is calculated between the state distributions of two policies, instead of only on the action probabilities, leading to a divergence augmentation formulation. Empirical experiments on Atari games show that in the data-scarce scenario where the reuse of off-policy data becomes necessary, our method can achieve better performance than other state-of-the-art deep reinforcement learning algorithms.
We explore a capability of evolution strategies to train an agent with its policy based on a transformer architecture in a reinforcement learning setting. We performed experiments using OpenAI's highly parallelizable evolution strategy to train Decision Transformer in Humanoid locomotion environment and in the environment of Atari games, testing the ability of this black-box optimization technique to train even such relatively large and complicated models (compared to those previously tested in the literature). We also proposed a method to aid the training by first pretraining the model before using the OpenAI-ES to train it further, and tested its effectiveness. The examined evolution strategy proved to be, in general, capable of achieving strong results and managed to obtain high-performing agents. Therefore, the pretraining was shown to be unnecessary; yet still, it helped us observe and formulate several further insights.
Recent advances in reinforcement learning have demonstrated the potential of quantum learning models based on parametrized quantum circuits as an alternative to deep learning models. On the one hand, these findings have shown the ultimate exponential speed-ups in learning that full-blown quantum models can offer in certain -- artificially constructed -- environments. On the other hand, they have demonstrated the ability of experimentally accessible PQCs to solve OpenAI Gym benchmarking tasks. However, it remains an open question whether these near-term QRL techniques can be successfully applied to more complex problems exhibiting high-dimensional observation spaces. In this work, we bridge this gap and present a hybrid model combining a PQC with classical feature encoding and post-processing layers that is capable of tackling Atari games. A classical model, subjected to architectural restrictions similar to those present in the hybrid model is constructed to serve as a reference. Our numerical investigation demonstrates that the proposed hybrid model is capable of solving the Pong environment and achieving scores comparable to the classical reference in Breakout. Furthermore, our findings shed light on important hyperparameter settings and design choices that impact the interplay of the quantum and classical components. This work contributes to the understanding of near-term quantum learning models and makes an important step towards their deployment in real-world RL scenarios.
This work analyses the disparity in performance between Decision Transformer (DT) and Decision Mamba (DM) in sequence modelling reinforcement learning tasks for different Atari games. The study first observed that DM generally outperformed DT in the games Breakout and Qbert, while DT performed better in more complicated games, such as Hero and Kung Fu Master. To understand these differences, we expanded the number of games to 12 and performed a comprehensive analysis of game characteristics, including action space complexity, visual complexity, average trajectory length, and average steps to the first non-zero reward. In order to further analyse the key factors that impact the disparity in performance between DT and DM, we employ various approaches, including quantifying visual complexity, random forest regression, correlation analysis, and action space simplification strategies. The results indicate that the performance gap between DT and DM is affected by the complex interaction of multiple factors, with the complexity of the action space and visual complexity (particularly evaluated by compression ratio) being the primary determining factors. DM performs well in environments with simple action and visual elements, while DT shows an advantage in games with higher action and visual complexity. Our findings contribute to a deeper understanding of how the game characteristics affect the performance difference in sequential modelling reinforcement learning, potentially guiding the development of future model design and applications for diverse and complex environments.
Training deep reinforcement learning (RL) agents necessitates overcoming the highly unstable nonconvex stochastic optimization inherent in the trial-and-error mechanism. To tackle this challenge, we propose a physics-inspired optimization algorithm called relativistic adaptive gradient descent (RAD), which enhances long-term training stability. By conceptualizing neural network (NN) training as the evolution of a conformal Hamiltonian system, we present a universal framework for transferring long-term stability from conformal symplectic integrators to iterative NN updating rules, where the choice of kinetic energy governs the dynamical properties of resulting optimization algorithms. By utilizing relativistic kinetic energy, RAD incorporates principles from special relativity and limits parameter updates below a finite speed, effectively mitigating abnormal gradient influences. Additionally, RAD models NN optimization as the evolution of a multi-particle system where each trainable parameter acts as an independent particle with an individual adaptive learning rate. We prove RAD's sublinear convergence under general nonconvex settings, where smaller gradient variance and larger batch sizes contribute to tighter convergence. Notably, RAD degrades to the well-known adaptive moment estimation (ADAM) algorithm when its speed coefficient is chosen as one and symplectic factor as a small positive value. Experimental results show RAD outperforming nine baseline optimizers with five RL algorithms across twelve environments, including standard benchmarks and challenging scenarios. Notably, RAD achieves up to a 155.1% performance improvement over ADAM in Atari games, showcasing its efficacy in stabilizing and accelerating RL training.
Agent learning from human interaction often relies on explicit signals, but implicit social cues, such as prosody in speech, could provide valuable information for more effective learning. This paper advocates for the integration of prosody as a teaching signal to enhance agent learning from human teachers. Through two exploratory studies--one examining voice feedback in an interactive reinforcement learning setup and the other analyzing restricted audio from human demonstrations in three Atari games--we demonstrate that prosody carries significant information about task dynamics. Our findings suggest that prosodic features, when coupled with explicit feedback, can enhance reinforcement learning outcomes. Moreover, we propose guidelines for prosody-sensitive algorithm design and discuss insights into teaching behavior. Our work underscores the potential of leveraging prosody as an implicit signal for more efficient agent learning, thus advancing human-agent interaction paradigms.
We introduce the Continuous Arcade Learning Environment (CALE), an extension of the well-known Arcade Learning Environment (ALE) [Bellemare et al., 2013]. The CALE uses the same underlying emulator of the Atari 2600 gaming system (Stella), but adds support for continuous actions. This enables the benchmarking and evaluation of continuous-control agents (such as PPO [Schulman et al., 2017] and SAC [Haarnoja et al., 2018]) and value-based agents (such as DQN [Mnih et al., 2015] and Rainbow [Hessel et al., 2018]) on the same environment suite. We provide a series of open questions and research directions that CALE enables, as well as initial baseline results using Soft Actor-Critic. CALE is available as part of the ALE athttps://github.com/Farama-Foundation/Arcade-Learning-Environment.
Mastering games is a hard task, as games can be extremely complex, and still fundamentally different in structure from one another. While the AlphaZero algorithm has demonstrated an impressive ability to learn the rules and strategy of a large variety of games, ranging from Go and Chess, to Atari games, its reliance on extensive computational resources and rigid Convolutional Neural Network (CNN) architecture limits its adaptability and scalability. A model trained to play on a $19\times 19$ Go board cannot be used to play on a smaller $13\times 13$ board, despite the similarity between the two Go variants. In this paper, we focus on Chess, and explore using a more generic Graph-based Representation of a game state, rather than a grid-based one, to introduce a more general architecture based on Graph Neural Networks (GNN). We also expand the classical Graph Attention Network (GAT) layer to incorporate edge-features, to naturally provide a generic policy output format. Our experiments, performed on smaller networks than the initial AlphaZero paper, show that this new architecture outperforms previous architectures with a similar number of parameters, being able to increase playing strength an order of magnitude faster. We also show that the model, when trained on a smaller $5\times 5$ variant of chess, is able to be quickly fine-tuned to play on regular $8\times 8$ chess, suggesting that this approach yields promising generalization abilities. Our code is available at https://github.com/akulen/AlphaGateau.
Playing two-player games using reinforcement learning and self-play can be challenging due to the complexity of two-player environments and the possible instability in the training process. We propose that a reinforcement learning algorithm can train more efficiently and achieve improved performance in a two-player game if it leverages the knowledge from the single-player version of the same game. This study examines the proposed idea in ten different Atari 2600 environments using the Atari 2600 RAM as the input state. We discuss the advantages of using transfer learning from a single-player training process over training in a two-player setting from scratch, and demonstrate our results in a few measures such as training time and average total reward. We also discuss a method of calculating RAM complexity and its relationship to performance.
Natural intelligence processes experience as a continuous stream, sensing, acting, and learning moment-by-moment in real time. Streaming learning, the modus operandi of classic reinforcement learning (RL) algorithms like Q-learning and TD, mimics natural learning by using the most recent sample without storing it. This approach is also ideal for resource-constrained, communication-limited, and privacy-sensitive applications. However, in deep RL, learners almost always use batch updates and replay buffers, making them computationally expensive and incompatible with streaming learning. Although the prevalence of batch deep RL is often attributed to its sample efficiency, a more critical reason for the absence of streaming deep RL is its frequent instability and failure to learn, which we refer to as stream barrier. This paper introduces the stream-x algorithms, the first class of deep RL algorithms to overcome stream barrier for both prediction and control and match sample efficiency of batch RL. Through experiments in Mujoco Gym, DM Control Suite, and Atari Games, we demonstrate stream barrier in existing algorithms and successful stable learning with our stream-x algorithms: stream Q, stream AC, and stream TD, achieving the best model-free performance in DM Control Dog environments. A set of common techniques underlies the stream-x algorithms, enabling their success with a single set of hyperparameters and allowing for easy extension to other algorithms, thereby reviving streaming RL.
A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization. We will release codes and model weights at https://github.com/CJReinforce/JOWA.