REVIEW 6 cited by
Stop Regressing: Training Value Functions via Classification for Scalable Deep RL
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost.
Forward citations
Cited by 6 Pith papers
-
WARP-RM: A Warp-Augmented Relative Progress Reward Model for Data Curation
WARP trains a reward model on time-warped successful demonstrations to produce frame-level progress estimates that upweight high-advantage chunks during behavior cloning, maintaining high success rates on suboptimal d...
-
SERNF: Sample-Efficient Real-World Dexterous Policy Fine-Tuning via Action-Chunked Critics and Normalizing Flows
SERNF fine-tunes dexterous manipulation policies on real hardware by pairing normalizing-flow policies with action-chunked critics and conservative off-policy RL.
-
ReBRAC-v2: The Return of the King
A fixed-recipe offline RL method combining normalizing-flow actors, categorical critics, staged training, and test-time refinement beats recent flow-based baselines by 22.5 points averaged over ten OGBench categories.
-
Relative Value Learning
A critic that learns antisymmetric value differences ∆(s_i,s_j)=V(s_i)−V(s_j) has a provably contracting Bellman operator and an unbiased advantage estimator, and PPO with this critic matches standard PPO on Atari.
-
RN-D: Discretized Categorical Actors for On-Policy Reinforcement Learning
A discretized categorical actor paired with a pre-LayerNorm residual MLP (RN-D) improves PPO's final performance and sample efficiency on continuous-control benchmarks.
-
Value Flows
Value Flows fits the full return distribution in RL with a flow-matching critic and reweights its learning objective by estimated return variance; the central theoretical guarantee does not follow from the stated equations.
Discussion (0). Continue with ORCID to comment.