As models grow larger and training them becomes expensive, it becomes increasingly important to scale training recipes not just to larger models and more data, but to do so in a compute-optimal manner that extracts maximal performance per unit of compute. While such scaling has been well studied for language modeling, reinforcement learning (RL) has received less attention in this regard. In this paper, we investigate compute scaling for online, value-based deep RL. These methods present two primary axes for compute allocation: model capacity and the updateto-data (UTD) ratio. Given a fixed compute budget, we ask: how should resources be partitioned across these axes to maximize data efficiency? Our analysis reveals a nuanced interplay between model size, batch size, and UTD. In particular, we identify a phenomenon we call TD-overfitting: increasing the batch quickly harms Q-function accuracy for small models, but this effect is absent in large models, enabling effective use of large batch size at scale. We provide a mental model for understanding this phenomenon and build guidelines for choosing batch size and UTD to optimize compute usage. Our findings provide a grounded starting point for compute-optimal scaling in deep RL, mirroring studies in supervised learning but adapted to TD learning.

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics, which are emphasized by higher UTD ratios. To address these, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, has been shown to prevent potential loss of plasticity, and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive performance across 25 challenging continuous control tasks on the DeepMind Control Suite and Myosuite benchmarks, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a simple yet robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

As models grow larger and training them becomes expensive, it becomes increasingly important to scale training recipes not just to larger models and more data, but to do so in a compute-optimal manner that extracts maximal performance per unit of compute. While such scaling has been well studied for language modeling, reinforcement learning (RL) has received less attention in this regard. In this paper, we investigate compute scaling for online, value-based deep RL. These methods present two primary axes for compute allocation: model capacity and the update-to-data (UTD) ratio. Given a fixed compute budget, we ask: how should resources be partitioned across these axes to maximize sample efficiency? Our analysis reveals a nuanced interplay between model size, batch size, and UTD. In particular, we identify a phenomenon we call TD-overfitting: increasing the batch quickly harms Q-function accuracy for small models, but this effect is absent in large models, enabling effective use of large batch size at scale. We provide a mental model for understanding this phenomenon and build guidelines for choosing batch size and UTD to optimize compute usage. Our findings provide a grounded starting point for compute-optimal scaling in deep RL, mirroring studies in supervised learning but adapted to TD learning.

Off-policy reinforcement learning in continuous control tasks depends critically on accurate $Q$-value estimates. Conservative aggregation over ensembles, such as taking the minimum, is commonly used to mitigate overestimation bias. However, these static rules are coarse, discard valuable information from the ensemble, and cannot adapt to task-specific needs or different learning regimes. We propose Directional Ensemble Aggregation (DEA), an aggregation method that adaptively combines $Q$-value estimates in actor-critic frameworks. DEA introduces two fully learnable directional parameters: one that modulates critic-side conservatism and another that guides actor-side policy exploration. Both parameters are learned using ensemble disagreement-weighted Bellman errors, which weight each sample solely by the direction of its Bellman error. This directional learning mechanism allows DEA to adjust conservatism and exploration in a data-driven way, adapting aggregation to both uncertainty levels and the phase of training. We evaluate DEA across continuous control benchmarks and learning regimes - from interactive to sample-efficient - and demonstrate its effectiveness over static ensemble strategies.

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics which are emphasized by higher UTDs, particularly Q-bias explosion and the growing magnitude of critic network weights. To address this, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, prevents potential loss of plasticity and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive or superior performance across a range of challenging tasks on the DeepMind control benchmark, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics, which are emphasized by higher UTD ratios. To address these, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, has been shown to prevent potential loss of plasticity and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive performance across 25 challenging continuous control tasks on the DeepMind Control Suite and Myosuite benchmarks, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a simple yet robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

Scaling data and compute is critical to the success of machine learning. However, scaling demands predictability: we want methods to not only perform well with more compute or data, but also have their performance be predictable from small-scale runs, without running the large-scale experiment. In this paper, we show that value-based off-policy RL methods are predictable despite community lore regarding their pathological behavior. First, we show that data and compute requirements to attain a given performance level lie on a Pareto frontier, controlled by the updates-to-data (UTD) ratio. By estimating this frontier, we can predict this data requirement when given more compute, and this compute requirement when given more data. Second, we determine the optimal allocation of a total resource budget across data and compute for a given performance and use it to determine hyperparameters that maximize performance for a given budget. Third, this scaling behavior is enabled by first estimating predictable relationships between hyperparameters, which is used to manage effects of overfitting and plasticity loss unique to RL. We validate our approach using three algorithms: SAC, BRO, and PQL on DeepMind Control, OpenAI gym, and IsaacGym, when extrapolating to higher levels of data, compute, budget, or performance.

A key challenge in Deep Reinforcement Learning is sample efficiency, especially in real-world applications where collecting environment interactions is expensive or risky. Recent off-policy algorithms improve sample efficiency by increasing the Update-To-Data (UTD) ratio and performing more gradient updates per environment interaction. While this improves sample efficiency, it significantly increases computational cost due to the higher number of gradient updates required. In this paper we propose a sample-efficient method to improve computational efficiency by separating training into distinct learning phases in order to exploit gradient updates more effectively. Our approach builds on top of the Dropout Q-Functions (DroQ) algorithm and alternates between an online, low UTD ratio training phase, and an offline stabilization phase. During the stabilization phase, we fine-tune the Q-functions without collecting new environment interactions. This process improves the effectiveness of the replay buffer and reduces computational overhead. Our experimental results on continuous control problems show that our method achieves results comparable to state-of-the-art, high UTD ratio algorithms while requiring 56\% fewer gradient updates and 50\% less training time than DroQ. Our approach offers an effective and computationally economical solution while maintaining the same sample efficiency as the more costly, high UTD ratio state-of-the-art.

We present a novel method aimed at enhancing the sample efficiency of ensemble Q learning. Our proposed approach integrates multi-head self-attention into the ensembled Q networks while bootstrapping the state-action pairs ingested by the ensemble. This not only results in performance improvements over the original REDQ (Chen et al. 2021) and its variant DroQ (Hi-raoka et al. 2022), thereby enhancing Q predictions, but also effectively reduces both the average normalized bias and standard deviation of normalized bias within Q-function ensembles. Importantly, our method also performs well even in scenarios with a low update-to-data (UTD) ratio. Notably, the implementation of our proposed method is straightforward, requiring minimal modifications to the base model.

The primacy bias in deep reinforcement learning (DRL), which refers to the agent's tendency to overfit early data and lose the ability to learn from new data, can significantly decrease the performance of DRL algorithms. Previous studies have shown that employing simple techniques, such as resetting the agent's parameters, can substantially alleviate the primacy bias. However, we observe that resetting the agent's parameters harms its performance in the context of model-based reinforcement learning (MBRL). In fact, on further investigation, we find that the primacy bias in MBRL differs from that in model-free RL. In this work, we focus on investigating the primacy bias in MBRL and propose world model resetting, which works in MBRL. We apply our method to two different MBRL algorithms, MBPO and DreamerV2. We validate the effectiveness of our method on multiple continuous control tasks on MuJoCo and DeepMind Control Suite, as well as discrete control tasks on Atari 100k benchmark. The results show that world model resetting can significantly alleviate the primacy bias in model-based setting and improve algorithm's performance. We also give a guide on how to perform world model resetting effectively.