Transfer in Reinforcement Learning via Shared Features

By George Dimitri Konidaris, Ilya Scheidwasser, Andrew Barto. 2012

117 citations recorded by the source corpus.

Subfields: Reinforcement Learning, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Reinforcement Learning

In Representation & Self-Supervised Learning

Citations in the corpus

Learning Parameterized Skills (2012, methodology): “[36, 40] introduced a method for learning reusable options by representing their policies not in the original state space, but in an agent-centered feature space.”
Learning to Act by Predicting the Future (2016, background): “Parameterized goals have been studied in the context of continuous motor skills such as throwing darts at a target (da Silva et al., 2012; Kober et al., 2012; Deisenroth et al., 2014).”
Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning (2017, background): “Typically these have the form of a meta controller and a set of lower-level controllers for subtasks (Sutton et al., 1999; Dietterich, 2000; Parr and Russell, 1997; Ghavamzadeh and Mahadevan, 2003; Konidaris et al., 2012; Konidaris and Barto, 2007).”
Meta Reinforcement Learning with Latent Variable Gaussian Processes (2018, background): “Much of the literature on meta and transfer learning in RL has focused on multi-task learning, i.e., cases where the system/robot is the same, but the task changes [16, 29, 3, 5, 20, 21, 24, 8, 12].”
Time-regularized interrupting options (2014, background): “Because of their flexibility options are often better suited for modeling complex problems than primitive actions (Stone et al., 2005; Konidaris et al., 2012).”
Multi-task policy search for robotics (2014, background): “For instance, in [16], a value function-based transfer learning approach is proposed that generalizes across tasks by finding a regression function mapping a task-augmented”
Autonomous Agents and Multi-agent Systems Manuscript No. Learning Potential Functions and Their Representations for Multi-task Reinforcement Learning (unknown): “Sim-ilartothedomain-relevantfeaturesdeﬁnedinthisarticleandtotheagentspaceofKonidaris andBarto[34],FrommbergerandWolter[22]deﬁne structure space as the feature space that retains the same semantics across tasks, and learn a structure space policy between tasks.”
Bisimulation Makes Analogies in Goal-Conditioned Reinforcement Learning (2022, unknown): “…to learn analogies, but enforces it through goal-conditioned bisimulation, making it a more general method for RL. Metrics have also been leveraged in RL for both MDPs and states for the goal of positive transfer (Carroll and Seppi, 2005; Fern´andez and Veloso, 2006; Konidaris et al., 2012).”
Portable Option Discovery for Automated Learning Transfer in Object-Oriented Markov Decision Processes (2015, background): “Other work considers reframing the domain into a new state space that is simpler to transfer [Konidaris and Barto, 2007; Konidaris et al., 2012].”
Learning Shared Representations in Multi-task Reinforcement Learning (2016, background): “A more recent study in (Konidaris et al., 2012), also employs the idea of a shared feature space, but both the learning procedure and the proposed way of transferring this tuned knowledge is very different from ours.”
Measuring the Distance Between Finite Markov Decision Processes (2016, background): “Reward-linked [17] indicates that the rewards are equal when reaching the same subgoals (or goals) in two MDPs, such as arriving at the exports in different mazes and eating pac-dots in different Pac-man Games [1].”