Reinforcement Learning
TarGF: Learning Target Gradient Field to Rearrange Objects without Explicit Goal Specification Mingdong Wu
Object Rearrangement is to move objects from an initial state to a goal state.
Tight High Probability Bounds for Linear Stochastic Approximation with Fixed Stepsize
However, in most cases, there is a consistent gap between these two types of analyses.
Residual Pathway Priors for Soft Equivariance Constraints Marc Finzi New York University Greg Benton New York University Andrew Gordon Wilson New York University
A disadvantage of hard coding these restrictions is that this prior knowledge may not match reality. A scene may have long range non-local interactions, rotation equivariance may be violated by a preferred camera angle, or a dynamical system may occasionally have discontinuous transitions. In particular, symmetries are delicate. A small perturbation like adding wind breaks the rotational symmetry of a pendulum, and bumpy or tilted terrain could break the translation symmetry for locomotion.
BYOL-Explore: Exploration by Bootstrapped Prediction
's intrinsic reward, whereas prior work could only get off the ground with human
Provable Benefit of Multitask Representation Learning in Reinforcement Learning
Our result demonstrates that multitask representation learning is provably more sample-efficient than learning each task individually, as long as the total number of tasks is above a certain threshold.