-
Notifications
You must be signed in to change notification settings - Fork 0
/
main-PROBO-benchmarking-graphene.R
137 lines (103 loc) · 5.49 KB
/
main-PROBO-benchmarking-graphene.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
# this file runs a simulation of n = 60 BO runs with a Budget
# of 90 evaluations and initial design of size 10 to benchmark PROBO
# against classical BO with several aquisition functions
# a detailed description of the experiments can be found in the
# paper in section 4
library(mlrMBO)
library(rgenoud)
library(BBmisc)
library(DiceKriging)
library(ggplot2)
library(randomForest)
# set seed
set.seed(293874)
# source graphene (kapton) data and respective RF as target function (including visualization)
source("data/make-kapton-rf.R") # time
# source files that integrate GLCB into mlrMBO
source("PROBO/makeMBOInfillCritGLCB.R")
source("PROBO/initCrit.InfillCritGLCB.R")
# globally only one target fun
fun <- kapton_fun_rf
# define sample size (number of BO runs per test function)
n <- 60
# define total Budget of evaluations per BO run
budget <- 90
# define initial design size
init_design <- 10
# list that is used inside foreach
results_list <- list()
# # OPTIONAL parallelize over functions WARNING only tested on linux
# cl <- parallel::makeCluster(number_of_funs)
# doParallel::registerDoParallel(cl)
# results_list <- foreach::foreach(i = 1:number_of_funs,
# .packages = c("mlrMBO", "smoof", "DiceKriging", "soobench")) %dopar% {
for (i in 1:1) { # allows for assessing more than one function or more than one initial sampling strategies
results_one_design <- list()
for (j in 1:n) {
print(j)
print(Sys.time())
## same design for all approaches, sampled anew each iteration of 1:n
design <- generateDesign(n = 10L, par.set = parameter_set, fun = lhs::randomLHS)
# set Control Argument of BO
ctrl <- makeMBOControl(propose.points = 1L)
# iters = budget
ctrl <- setMBOControlTermination(ctrl, iters = budget)
#set control globally
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritCB(cb.lambda = 1), opt = "focussearch",
opt.focussearch.points = 1000, opt.focussearch.maxit = 5)
## classic bo with classic GP with powexp kernel
lrn_classic <- makeLearner("regr.km", covtype = "powexp", predict.type = "se", optim.method = "gen",
control = list(trace = FALSE), config = list(on.par.without.desc = "warn"))
# ensure numerical stability in km {DiceKriging} cf. github issue and recommendation by Bernd Bischl
y = fun(design)
Nuggets = 1e-8*var(y)
lrn_classic = setHyperPars(learner = lrn_classic, nugget=Nuggets)
#start classic bo with LCB and lambda = 1
cl_bo_res_lcb <- mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# bo with adaptive LCB
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritAdaCB())
cl_bo_res_alcb <-mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# bo with EI
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritEI())
cl_bo_res_ei <-mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# bo with augmented EI
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritAEI())
cl_bo_res_aei <-mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# bo with infill crit = expected quantile improvement
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritEQI())
cl_bo_res_eqi <-mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# bo with infill crit = se
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritStandardError())
cl_bo_res_se <-mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
# plug-in-bo with Generalized Lower Confidence Bound
# set Control Argument of BO
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritGLCB(imprecision = 50, cb.rho = 1, base_kernel = "powexp"))
glcb_1_50_bo_res <- mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritGLCB(imprecision = 100, cb.rho = 1, base_kernel = "powexp"))
glcb_1_100_bo_res <- mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
ctrl <- setMBOControlInfill(ctrl, crit = makeMBOInfillCritGLCB(imprecision = 100, cb.rho = 10, base_kernel = "powexp"))
glcb_10_100_bo_res <- mbo(fun, design, ctrl, learner = lrn_classic, show.info = FALSE)
opt_paths_raw <- list(getOptPathY(cl_bo_res_lcb[["opt.path"]]),
getOptPathY(cl_bo_res_alcb[["opt.path"]]),
getOptPathY(cl_bo_res_ei[["opt.path"]]),
getOptPathY(cl_bo_res_aei[["opt.path"]]),
getOptPathY(cl_bo_res_eqi[["opt.path"]]),
getOptPathY(cl_bo_res_se[["opt.path"]]),
getOptPathY(glcb_1_50_bo_res[["opt.path"]]),
getOptPathY(glcb_1_100_bo_res[["opt.path"]]),
getOptPathY(glcb_10_100_bo_res[["opt.path"]]))
# ATTENTION only for minimization (no problem here, since we restrict our tests to it)
opt_paths <- lapply(opt_paths_raw, function(proposals){
for (o in 2:length(proposals)) {
if(proposals[o] < proposals[o - 1])
proposals[o] = proposals[o - 1]
}
proposals
})
results_one_design[[j]] <- opt_paths
}
results_list[[i]] <- results_one_design
print(Sys.time())
}
# save results in repo so that it can be accessed and visualized later
save(results_list, file = paste(getwd(),"/benchmarking/results/PROBO-results-graphene-new" ,sep=""))