Hadamard Representation: Scaffolding Performance Across Model-free RL
Researchers have introduced a new architectural component called the Hadamard Representation (HR) to address performance degradation in deep reinforcement learning agents. HR replaces a standard hidden layer with an element-wise product of two independently parameterized layers. This method aims to prevent neurons from becoming dormant and to increase effective rank, thereby capturing richer feature interactions. Evaluations across multiple algorithms and domains demonstrated consistent performance improvements over baseline models without requiring hyperparameter tuning. AI
IMPACT Introduces a novel architectural component that enhances the stability and performance of deep reinforcement learning agents.