SMDP homomorphisms: an algebraic approach to abstraction in semi-Markov decision processes

By Balaraman Ravindran, Andrew Barto. 2003

59 citations recorded by the source corpus.

Citations in the corpus

• RUDDER: Return Decomposition for Delayed Rewards (2018, neutral): “between MDPs maps a MDP to another one with corresponding reward and transitions probabilities. Ravindran and Barto have shown that solving the original MDP can be done by solving a homomorphic image [99]. Therefore, Ravindran and Barto have also shown that state-enriched MDPs preserve the optimal a”
• Value Preserving State-Action Abstractions (2020, unknown): “With a smaller state space, learning algorithms can learn with less computation, space, and even samples (Dearden & Boutilier, 1997; Dietterich, 2000; Ravindran, 2003; Jong & Stone, 2005; Odalric-Ambrym et al., 2013; Hostetler et al., 2014; Jiang et al., 2015).”
• Hierarchical Approaches (2012, unknown): “Ravindran and Barto (2003) extend this form of model minimisation to state-action symmetry couched in the algebraic formalism of morphisms.”
• Autonomous identification, categorization and generalization of policies based on task type (2011, unknown): “Ravindran and Barto [10][11] develop a formal framework based on MDP homomorphism that extracts a smaller state space from the original by exploiting its symmetries and redundancies.”