Squeezing More from the Stream : Learning Representation Online for Streaming Reinforcement Learning

In streaming Reinforcement Learning (RL), transitions are observed and discarded immediately after a single update. While this minimizes resource usage for on-device applications, it makes agents notoriously sample-inefficient, since value-based losses alone struggle to extract meaningful representations from transient data. We propose extending Self-Predictive Representations (SPR) to the streaming pipeline to maximize the utility of every observed frame. However, due to the highly correlated samples induced by the streaming regime, naively applying this auxiliary loss results in training instabilities. Thus, we introduce orthogonal gradient updates relative to the momentum target and resolve gradient conflicts arising from streaming-specific optimizers. Validated across the Atari, MinAtar, and Octax suites, our approach systematically outperforms existing streaming baselines. Latent-space analysis, including t-SNE visualizations and effective-rank measurements, confirms that our method learns significantly richer representations, bridging the performance gap caused by the absence of a replay buffer, while remaining efficient enough to train on just a few CPU cores.

Many learning systems improve by storing past experiences and reviewing them many times, much like a student studying old notes. But this is not always possible: robots, phones, or other small devices may have limited memory, limited computing power, or privacy constraints that prevent them from keeping large records of what they observe. This paper studies “streaming” reinforcement learning, where an agent must learn from each experience immediately and then discard it. Current methods can make this kind of learning stable, but they often waste useful information because each experience is used only once. We propose a way for the agent to learn more from every observation by adding an extra training signal that asks it to predict how its internal understanding of the world should change over the next few steps. Because streaming data is highly repetitive and closely connected over time, we also adjust the learning updates so this extra signal helps rather than destabilizes training. Across Atari, MinAtar, and Octax games, this approach learns faster, performs better, and builds richer internal representations while remaining efficient enough for CPU-based training.