A research method called TERC promises to make reinforcement learning agents faster by teaching them to ignore what doesn't matter.
Researchers introduced the Transfer Entropy Redundancy Criterion, or TERC, a technique that identifies which observable state variables actually influence an agent's decisions during training. The method uses information theory to measure whether a given variable transfers any entropy to the agent's actions — if it doesn't, TERC removes it from the state representation. The result is a leaner input space that the researchers say can cut inference time by up to 2.6 times. They tested the approach across tabular Q-learning, Actor-Critic, and Proximal Policy Optimization — three of the most common RL algorithm families — and across multiple environments.
State variable bloat is a quiet tax on every RL deployment. Agents trained on high-dimensional observations often carry dead weight: inputs that looked relevant at design time but contribute nothing to policy decisions. TERC offers a principled, policy-dependent way to find and drop that weight without re-architecting the underlying algorithm.
The approach won't rewrite production pipelines overnight — TERC still requires a training run to identify which variables are expendable, which adds upfront cost. But for teams retraining agents frequently, or running inference at scale, a 2.6x speedup from tidying up the state space is the kind of unglamorous win that compounds.