Tunable algorithms fortransient follow-up
Tim Staley
TKP Meeting, Manchester, Sept 2014
WWW: 4pisky.org , timstaley.co.uk
Context Theory Implementation Future work Fin
Aim of this talk
A basic, intuitive understanding of
information content
and how this can be used tooptimize / automate decision
making, a.k.a.
Bayesian decision theory
Context Theory Implementation Future work Fin
Outline
Context
Theory
Implementation
Future work
Fin
Context Theory Implementation Future work Fin
A blueprint for automatedfollow-up
Context Theory Implementation Future work Fin
Outline
Context
Theory
Implementation
Future work
Fin
Context Theory Implementation Future work Fin
6 4 2 0 2 4 6Epoch
0.0
0.2
0.4
0.6
0.8
1.0
Rel
ativ
e flu
x
stablelogisticnull
Intrinsic lightcurves
Context Theory Implementation Future work Fin
6 4 2 0 2 4 6Epoch
0.5
0.0
0.5
1.0
1.5
Rel
ativ
e flu
x
True valueNoisy samples
Sampling with noise
Context Theory Implementation Future work Fin
6 4 2 0 2 4 6Epoch
0.5
0.0
0.5
1.0
1.5
Rel
ativ
e flu
xSampling with noise
Context Theory Implementation Future work Fin
6 4 2 0 2 4 6Epoch
0.5
0.0
0.5
1.0
1.5
Rel
ativ
e flu
xSampling with noise
Context Theory Implementation Future work Fin
0 10.5
0.0
0.5
1.0
1.5R
elat
ive
flux
T=-5.0
0 1
T=-4.0
0 1PDF value
T=-3.0
0 1
T=-2.0
0 1
T=-1.0
0 1
T=0.0
0 1
T=1.0
0 1
T=2.0
stablelogisticnull
Class PDF at each epoch
Context Theory Implementation Future work Fin
PDF value0.5
0.0
0.5
1.0
1.5
Rel
ativ
e flu
x
T=-5.0 T=-4.0 T=-3.0 T=-2.0 T=-1.0 T=0.0 T=1.0 T=2.0
stablelogisticnull
5 4 3 2 1 0 1 2Epoch
0.500.450.400.350.300.25
FoM Information content
Evaluating each epoch
Context Theory Implementation Future work Fin
Confusion matrices
Context Theory Implementation Future work Fin
Confusion matrices
Context Theory Implementation Future work Fin
−4 −2 0 2 4Time
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Featu
re v
alu
e
stable
logistic
Instrinsic lightcurves - ensemble
Context Theory Implementation Future work Fin
Outline
Context
Theory
Implementation
Future work
Fin
Context Theory Implementation Future work Fin
Required knowledge /user-inputs
É Transient rate priors.
É Transient lightcurve ensemble models.
É Telescope / noise models.
É Follow-up prioritization weightings.
Context Theory Implementation Future work Fin
Required knowledge /user-inputs
É Transient rate priors.
É Transient lightcurve ensemble models.
É Telescope / noise models.
É Follow-up prioritization weightings.
Context Theory Implementation Future work Fin
Required knowledge /user-inputs
É Transient rate priors.
É Transient lightcurve ensemble models.
É Telescope / noise models.
É Follow-up prioritization weightings.
Context Theory Implementation Future work Fin
Required knowledge /user-inputs
É Transient rate priors.
É Transient lightcurve ensemble models.
É Telescope / noise models.
É Follow-up prioritization weightings.
Context Theory Implementation Future work Fin
Required softwarecomponents
É Efficient lightcurve generation library.
É MCMC data fitting models androutines.
É Statistical routines for calculatingconfusion matrices.
É Observation schedule optimizationengine.
Context Theory Implementation Future work Fin
Required softwarecomponents
É Efficient lightcurve generation library.
É MCMC data fitting models androutines.
É Statistical routines for calculatingconfusion matrices.
É Observation schedule optimizationengine.
Context Theory Implementation Future work Fin
Required softwarecomponents
É Efficient lightcurve generation library.
É MCMC data fitting models androutines.
É Statistical routines for calculatingconfusion matrices.
É Observation schedule optimizationengine.
Context Theory Implementation Future work Fin
Required softwarecomponents
É Efficient lightcurve generation library.
É MCMC data fitting models androutines.
É Statistical routines for calculatingconfusion matrices.
É Observation schedule optimizationengine.
Context Theory Implementation Future work Fin
Required components
Context Theory Implementation Future work Fin
Outline
Context
Theory
Implementation
Future work
Fin
Context Theory Implementation Future work Fin
What’s next?
É Finish bolting components together.
É Run simulations, test in more realisticscenarios.
É Interfacing with optimizer / scheduler.
Context Theory Implementation Future work Fin
Longer term
É Variational Bayes?
É Gaussian processes?
Context Theory Implementation Future work Fin
Outline
Context
Theory
Implementation
Future work
Fin
Context Theory Implementation Future work Fin
Summary
É Information content is just a penaltyfunction for scoring predictedobservations.
É Using it to decide when to observe isapplied Bayesian decision theory.
É But doing this for real requires anumber of non-trivial softwarecomponents.
É Nearly ready for testing!