At 51 mins 42 there is a general explanation of algorithms which transform a given point selected at random, with the aim of being able to systematically explore the "quality neighbourhood" for points of higher quality.
Some of the statements he makes seem quite strong, such as "there is no better way to choose initial points than at random." But I feel that needs justification, if not only by qualification, because it seems obvious at first sight, to me, that one could choose different distributions and some would be more densely populated by higher quality samples than others. Evidence supporting this comes from the fact that when PRNG's are used to choose random points, they sometimes hit resonance points which have a lower density (zero) of higher quality points. So that suggests that one could try to better understand the conditions of resonance and try to construct PRNGs that are geared to the problem space and which resonate with "rich seams". I have some ideas of things which would be good fun to try. See for example my comments here.
Some of the statements he makes seem quite strong, such as "there is no better way to choose initial points than at random." But I feel that needs justification, if not only by qualification, because it seems obvious at first sight, to me, that one could choose different distributions and some would be more densely populated by higher quality samples than others. Evidence supporting this comes from the fact that when PRNG's are used to choose random points, they sometimes hit resonance points which have a lower density (zero) of higher quality points. So that suggests that one could try to better understand the conditions of resonance and try to construct PRNGs that are geared to the problem space and which resonate with "rich seams". I have some ideas of things which would be good fun to try. See for example my comments here.
No comments:
Post a Comment