Initial Conditions Pool Guide#
A Initial Conditions Pool (ICPool) inherits from BaseICPool
and manages a queue of InitialConditions.
An InitialCondition is
a tuple of BaseComponents combined with a unique, informative
label stored as a string and is
used to initialise a BaseMixture object.
The ICPool attempts to keep its queue populated with reasonable InitialConditions. It does this by generating a new generation of InitialConditions each time the queue becomes empty. Precisely how the new generation is generated depends on the specific implementation.
The ICPool expects the result of a fit to a given InitialCondition to be
registered back to it with register_result()
along with the fit’s score.
The ICPool can then use the scores to determine the best fit, and use that to
generate the next generation.
See the source code for Driver for a full example of
how to utilise an ICPool.
Implemented ICPools#
SimpleICPool#
A simple implementation SimpleICPool identifies the best mixture with \(k\)
components and uses this to generate a set of \(k+1\)-component initial conditions. This is the algorithm used in Chronostar Paper I.
The set of \(k+1\)-component initial conditions is formed thusly:
for each component (named the target component) in the best fitting \(k\)-component mixture, make
a new initial condition. This new initial condition has an identical set of components as the \(k\)-component mixture, except the target component.
The target component is split() into
two similar, overlapping components that when combined more or less describe the target component.
This is a safe approach that guarantees the discovery of the best Mixture for a given number of components. However, it involves a lot of repeated computation, since in order to go from \(k\) components to \(k+1\) compoents, \(k\) fits must be performed. One can see that this results in multiple splitting of the majority of the components. Indeed this algorithm scales quadratically with the number of components.
GreedyCycleICP#
A more efficient approach is the GreedyCycleICP. It is named the GreedyCycleICP because each time it finds a component that improves the fit’s
score, this ICPool immediately incorporates it into its best fit.
This approach is more efficient, because fewer fits are needed before a new component is found. This approach can miss the best solution, however, by prematurely accepting a component in one region of the dataspace, when a different region would have benefitted more, i.e. the solution gets stuck in a local minimum/maximum.