runBayesFit: Running the BayesFit optimisation

In jaegea/MEIGOR: MEIGOR - MEtaheuristics for bIoinformatics Global Optimization

Running the BayesFit optimisation

Description

"runBayesFit" defines the prior function and runs the BayesFit estimation

Usage

runBayesFit(opts)

Arguments

opts

list with entries as explained below. Options set – defines the problem and sets some parameters to control the MCMC algorithm. model: List of model parameters - to estimate. The parameter objects must each have a 'value' attribute containing the parameter's numerical value. estimate_params: list. List of parameters to estimate, all of which must also be listed in 'options$model$parameters'. initial_values: list of float, optional. Starting values for parameters to estimate. If omitted, will use the nominal values from 'options$model$parameters' step_fn: callable f(output), optional. User callback, called on every MCMC iteration. likelihood_fn: callable f(output, position). User likelihood function. prior_fn: callable f(output, position), optional. User prior function. If omitted, a flat prior will be used. nsteps: int. Number of MCMC iterations to perform. use_hessian: logical, optional. Wheter to use the Hessian to guide the walk. Defaults to FALSE. rtol: float or list of float, optional. Relative tolerance for ode solver. atol: float or list of float, optional. Absolute tolerance for ode solver. norm_step_size: float, optional. MCMC step size. Defaults to a reasonable value. hessian_period: int, optional. Number of MCMC steps between Hessian recalculations. Defaults to a reasonable but fairly large value, as Hessian calculation is expensive. hessian_scale: float, optional. Scaling factor used in generating Hessian-guided steps. Defaults to a reasonable value. sigma_adj_interval: int, optional. How often to adjust 'output$sig_value' while annealing to meet 'accept_rate_target'. Defaults to a reasonable value. anneal_length: int, optional. Length of initial "burn-in" annealing period. Defaults to 10 'nsteps', or if 'use_hessian' is TRUE, to 'hessian_period' (i.e. anneal until first hessian is calculated) T_init: float, optional. Initial temperature for annealing. Defaults to a resonable value. accept_rate_target: float, optional. Desired acceptance rate during annealing. Defaults to a reasonable value. See also 'sigma_adj_interval' above. sigma_max: float, optional. Maximum value for 'output$sig_value'. Defaults to a resonable value. sigma_min: float, optional. Minimum value for 'output$sig_value'. Defaults to a resonable value. sigma_step: float, optional. Increment for 'output$sig_value' adjustments. Defaults to a resonable value. To eliminate adaptive step size, set sigma_step to 1. thermo_temp: float in the range [0,1], optional. Temperature for thermodynamic integration support. Used to scale likelihood when calculating the posterior value. Defaults to 1, i.e. no effect.

Value

The output after the optimisation is finished - a list with entries as explained in 'Arguments'.

Examples

data("simpleExample", package="MEIGOR")
initial_pars = createLBodeContPars(model, LB_n=1, LB_k=0.1, LB_tau=0.01, UB_n=5, UB_k=0.9, UB_tau=10, random=TRUE)
simData = plotLBodeFitness(cnolist, model, initial_pars, reltol=1e-05, atol=1e-03, maxStepSize=0.01)

f_bayesFit <- function(position, params=initial_pars, exp_var=opts$exp_var) {
	# convert from log
	params$parValues = 10^position
	ysim = getLBodeDataSim(cnolist=cnolist, model=model,
	ode_parameters=params)
	data_as_vec = unlist(cnolist$valueSignals)
	sim_as_vec = unlist(ysim)
	# set nan (NAs) to 0
	sim_as_vec[is.na(sim_as_vec)] = 0
	sim_as_vec[is.nan(sim_as_vec)]= 0
	return(sum((data_as_vec-sim_as_vec)^2/(2*exp_var^2)))
}

prior_mean = log10(initial_pars$parValues)
prior_var = 10

opts <- list("model"=NULL, "estimate_params"=NULL,"initial_values"=NULL,
"tspan"=NULL, "step_fn"=NULL, "likelihood_fn"=NULL,
"prior_fn"=NULL, "nsteps"=NULL, "use_hessian"=FALSE,
"rtol"=NULL, "atol"=NULL, "norm_step_size"=0.75,
"hessian_period"=25000, "hessian_scale"=0.085,
"sigma_adj_interval"=NULL, "anneal_length"=NULL,
"T_init"=10, "accept_rate_target"=0.3, "sigma_max"=1,
"sigma_min"=0.25, "sigma_step"=0.125, "thermo_temp"=1, "seed"=NULL)
opts$nsteps = 2000
opts$likelihood_fn = f_bayesFit
opts$use_hessian = TRUE
opts$hessian_period = opts$nsteps/10
opts$model = list(parameters=list(name=initial_pars$parNames,
value=initial_pars$parValues))
opts$estimate_params = initial_pars$parValues
opts$exp_var = 0.01

res = runBayesFit(opts)

initial_pars$parValues = cur_params(output=res, options=opts)

jaegea/MEIGOR documentation built on April 8, 2024, 9:36 a.m.