search

tapios / risk-networks

Code for risk networks: a blend of compartmental models, graphs, data assimilation and semi-supervised learning
Other
2 stars 2 forks source link

Assimilate negative information of hospital & death records #107

Closed odunbar closed 4 years ago

odunbar commented 4 years ago

Previously we were assimilating only the positive records. That is, if node 342 enters hospital at some time window t we would assimilate the risk <H_{342}(t)> = 1. Now we also assimilate the risk of <H_i(t)>=0 for all i !=342. Note this is a more expensive task, as it always requires the assimilation of the state H (and for death records D) for all nodes.

We demonstrate the use case in the simple_epidemic_with_da_health_and_death_records.py. We create observations negative_health_record and negative_death_record, by setting a bool_type=False.

We set the risk of - for example H - for of every node with statuses !='H' to a small value near zero. In practice this must be much smaller than we have set previously, as it is applied to many more nodes in the network (and thus the implication on the whole population is more significant).

Co-authored (in spirit) with @agarbuno

odunbar commented 4 years ago

da_ric_tprobs_posdrec_negdrec

Here is setting (near) perfect states to 1 or 0 with error tolerance 1e-6. I have reduced this to 1e-10 for the next run to get closer agreement.

Seemingly this is as good as it can get: da_ric_tprobs_posdrec_negdrec