Below is a list of all relevant publications authored by Robotics Forum members.

Citation

BibTex format

@inproceedings{Rakicevic:2021:10.1145/3449639.3459320,
author = {Rakicevic, N and Cully, A and Kormushev, P},
doi = {10.1145/3449639.3459320},
title = {Policy manifold search: exploring the manifold hypothesis for diversity-based neuroevolution},
url = {http://dx.doi.org/10.1145/3449639.3459320},
year = {2021}
}

RIS format (EndNote, RefMan)

TY  - CPAPER
AB - Neuroevolution is an alternative to gradient-based optimisation that has thepotential to avoid local minima and allows parallelisation. The main limitingfactor is that usually it does not scale well with parameter spacedimensionality. Inspired by recent work examining neural network intrinsicdimension and loss landscapes, we hypothesise that there exists alow-dimensional manifold, embedded in the policy network parameter space,around which a high-density of diverse and useful policies are located. Thispaper proposes a novel method for diversity-based policy search viaNeuroevolution, that leverages learned representations of the policy networkparameters, by performing policy search in this learned representation space.Our method relies on the Quality-Diversity (QD) framework which provides aprincipled approach to policy search, and maintains a collection of diversepolicies, used as a dataset for learning policy representations. Further, weuse the Jacobian of the inverse-mapping function to guide the search in therepresentation space. This ensures that the generated samples remain in thehigh-density regions, after mapping back to the original space. Finally, weevaluate our contributions on four continuous-control tasks in simulatedenvironments, and compare to diversity-based baselines.
AU - Rakicevic,N
AU - Cully,A
AU - Kormushev,P
DO - 10.1145/3449639.3459320
PY - 2021///
TI - Policy manifold search: exploring the manifold hypothesis for diversity-based neuroevolution
UR - http://dx.doi.org/10.1145/3449639.3459320
UR - http://arxiv.org/abs/2104.13424v1
UR - http://hdl.handle.net/10044/1/88718
ER -