Page 385 - Special Topic Session (STS) - Volume 2
P. 385
STS506 L. Leticia R. et al.
surrogate-ABC-MCMC took less than 2 minutes to obtain the posterior
sample.
This ABC-MCMC modification becomes highly attractive not only because
it is fast to compute, but because it only requires partial knowledge of the
network, that is its first and second moment of its degree.
3.3 Assessing surrogate model B: “RNN”
The proposed RNN architecture allows us to predict the dynamics of
epidemic outbreaks on contact networks with similar degree distributions (in
the experiments depicted in Figure 3 we used networks with Poisson
generated degree sequences). Although this model requires a considerable
amount of time to train (approximately 5 minutes for two sets of simulations),
it offers fast prediction as it took an average of 38.5ms per simulation on a i7-
8650U processor.
Figure 3: Predicted outbreaks with our augmented RNN architecture.
4. Discussion and Conclusion
We base the likelihood-free inference ABC-MCMC on more realistic
surveillance–like information, that report only the new infective in intervals of
time. We emphasize on the inference for the SIR parameters and , but using
this same ABC methods, we can do the inference of any epidemic model from
which we can obtain pseudo-observations. The ABC-MCMC can be very
computer expensive if it uses the agent-based simulation, but in this case,
many other characteristics can be inferred, such as the degree of the initial
infectious cases, the most exposed community, etc. This approach also allows
doing inference for non-Markovian epidemic models.
We explore the use of surrogate models to run more efficient ABC-MCMC.
For the specific case of SIR (or SEIR) models, we propose harnessing model (1)
but other alternatives must be used when the model has infectious (latent)
periods are not exponentially distributed. In these more general scenarios, we
propose some models that allows for fast forecasting, while they can
incorporate the network topological features that are relevant for the outbreak
374 | I S I W S C 2 0 1 9