notesum.ai

Published at November 7

Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity

cs.LG

cs.AI

cs.RO

stat.ML

Released Date: November 7, 2024

Authors: Robby Costales¹, Stefanos Nikolaidis¹

Aff.: ¹Department of Computer Science, University of Southern California

Arxiv: http://arxiv.org/abs/2411.04466v1

Refer to caption

Name	Abbr.	Description
ManhattanToOptimal	mto	Average Manhattan distance between all stones (across all trials) to the optimal state.
StoneToStoneDistance	stsd	Average Euclidean distance between all pairs of stones (across all trials).
GraphNumBottlenecks	gnb	The number of bottlenecks in the graph topology.
LatentStateDiversity	lsd	The ’diversity’ of the latent stone states (across all trials). Diversity is calculated as the standard deviation of each latent state coordinate across all stones.
ParityFirstPotion	pfp	First potion location (first trial, first potion), as a parity measure.
ParityFirstStone	pfs	First stone location (first trial, first stone), as a parity measure.
PotionEffectDiversity	ped	The ’diversity’ of the potion effects (across all trials). Diversity is calculated as the standard deviation of each potion effect coordinate across all potions.
PotionPermutation	pp	Potion permutation.
PotionReflection	pr	Potion reflection.
StoneReflection	sre	Stone reflection.
StoneRotation	sro	Stone rotation.
StoneToStoneDistanceVariance	stsdv	Variance of the distances between stones (across all trials).