This work proposes a new methodology for discovering new species, when observations are sampled from different populations. Using a metaphor, we imagine J populations of animals to be available and we can sequentially choose from which of these populations to collect further samples. Both labels and fre- quencies of these species are unknown a priori. At each time step, the proposed strategy suggests where to collect the next observation in order to maximize the number of total species observed. This strategy is based on a joint use of the Hier- archical Pitman-Yor process, to estimate the unknown distributions of animals, and of Thompson Sampling for the sequential allocation problem. Performances of the algorithm are compared to those of other three strategies through simulations.
Thompson sampling for species discovery
FAVARO, STEFANO;
2016-01-01
Abstract
This work proposes a new methodology for discovering new species, when observations are sampled from different populations. Using a metaphor, we imagine J populations of animals to be available and we can sequentially choose from which of these populations to collect further samples. Both labels and fre- quencies of these species are unknown a priori. At each time step, the proposed strategy suggests where to collect the next observation in order to maximize the number of total species observed. This strategy is based on a joint use of the Hier- archical Pitman-Yor process, to estimate the unknown distributions of animals, and of Thompson Sampling for the sequential allocation problem. Performances of the algorithm are compared to those of other three strategies through simulations.File | Dimensione | Formato | |
---|---|---|---|
bandit.pdf
Accesso aperto
Tipo di file:
POSTPRINT (VERSIONE FINALE DELL’AUTORE)
Dimensione
223.46 kB
Formato
Adobe PDF
|
223.46 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.