Using Forwards-Backwards Models to Approximate MDP Homomorphisms
This addresses sample efficiency in reinforcement learning for agents, though it is incremental as it builds on existing MDP homomorphism concepts.
The paper tackles the problem of inefficient reinforcement learning by proposing a method to automatically construct MDP homomorphisms using learned dynamics models, which reduces the state-action space size by up to the original action space cardinality. In experiments on MinAtar games, they achieved an almost 4x improvement over a baseline in low-sample settings.
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 optimizers.