Using Forwards-backwards Models To Approximate MDP Homomorphisms

Abstract

Reinforcement learning agents must painstakingly learn through trial and error what sets of state-action pairs are value equivalent -- requiring an often prohibitively large amount of environment experience. MDP homomorphisms have been proposed that reduce the MDP of an environment to an abstract MDP, enabling better sample efficiency. Consequently, impressive improvements have been achieved when a suitable homomorphism can be constructed a priori -- usually by exploiting a practitioner's knowledge of environment symmetries. We propose a novel approach to constructing homomorphisms in discrete action spaces, which uses a learnt model of environment dynamics to infer which state-action pairs lead to the same state -- which can reduce the size of the state-action space by a factor as large as the cardinality of the original action space. In MinAtar, we report an almost 4x improvement over a value-based off-policy baseline in the low sample limit, when averaging over all games and optimiz

Stats

• citations0
• S2 citations—
• github stars0
• HF likes0
• heat score0.00
• arxiv keymavorparker2022using

Related papers

• Low-dimensional State And Action Representation Learning With MDP Homomorphism Metrics (2021)0.00
• Policy Gradient Methods In The Presence Of Symmetries And State Abstractions (2023)0.00
• Multi-agent MDP Homomorphic Networks (2021)0.00
• Deepmdp: Learning Continuous Latent Space Models For Representation Learning (2019)0.00
• Metric Policy Representations For Opponent Modeling (2021)0.00
• Backplay: "man Muss Immer Umkehren" (2018)0.00
• Mimicking Better By Matching The Approximate Action Distribution (2023)0.00
• Efficient Model-based Multi-agent Reinforcement Learning Via Optimistic Equilibrium Computation (2022)0.00