The Skill-Action Architecture: Learning Abstract Action Embeddings for Reinforcement Learning

The option framework, one of the most promising Hierarchical Reinforcement Learning (HRL) frameworks, is developed based on the Semi-Markov Decision Problem (SMDP) and employs a triple formulation of the option (i.e., an action policy, a termination probability, and an initiation set). These design choices, however, mean that the option framework: 1) has low sample efficiency, 2) cannot use more stable Markov Decision Problem (MDP) based learning algorithms, 3) represents abstract actions implicitly, and 4) is expensive to scale up. To overcome these problems, here we propose a simple yet effective MDP implementation of the option framework: the Skill-Action (SA) architecture. Derived from a novel discovery that the SMDP option framework has an MDP equivalence, SA hierarchically extracts skills (abstract actions) from primary actions and explicitly encodes these knowledge into skill context vectors (embedding vectors). Although SA is MDP formulated, skills can still be temporally extended by applying the attention mechanism to skill context vectors. Unlike the option framework, which requires $M$ action policies for $M$ skills, SA's action policy only needs one decoder to decode skill context vectors into primary actions. Under this formulation, SA can be optimized with any MDP based policy gradient algorithm. Moreover, it is sample efficient, cheap to scale up, and theoretically proven to have lower variance. Our empirical studies on challenging infinite horizon robot simulation games demonstrate that SA not only outperforms all baselines by a large margin, but also exhibits smaller variance, faster convergence, and good interpretability. A potential impact of SA is to pave the way for a large scale pre-training architecture in the reinforcement learning area.