hxd1011 - 7 months ago 22

R Question

Suppose I have an optimization problem to solve using

`R`

`optimx`

`f<-function(x){`

return(sum(abs(x)))

}

gr<-function(x){

return(sign(x))

}

opt=optimx::optimx(runif(2),f,gr,method="BFGS")

The goal is trying to make such plot:

I think we can manually do it with Gradient Decent with following code, but how I can I do it in

`optimx`

`x=c(0.5,1.5)`

alpha=0.1

max_iter=20

x_trace=matrix(rep(0,max_iter*2),ncol=2)

for (i in 1:max_iter){

x=x-alpha*gr(x)

x_trace[i,]=x

}

f_trace=apply(x_trace,1,f)

Answer

Create a side effect:

```
f<-function(x){
.GlobalEnv$i <- get("i", envir = .GlobalEnv) + 1
.GlobalEnv$log[get("i", envir = .GlobalEnv),] <- x
return(sum(abs(x)))
}
gr<-function(x){
return(sign(x))
}
library(optimx)
i <- 0
log <- matrix(numeric(100 * 2), ncol = 2)
opt <- optimx(c(0.8, -0.9),f,gr,method="BFGS")
log <- log[seq_len(i), ]
plot(log, type = "l", xlim = c(-2, 2), ylim = c(-1.2, 1.2))
```

Note that this includes all function calls, even those where the algorithm rejects the result and retries. `control = list(trace = TRUE, REPORT = 1)`

lets `optimx`

print the function values for accepted tries and you could `capture.output`

this and use it to get only the parameters of these from `log`

.

It would be better to change `optimx`

to return all accepted attempts, but I'm not going to invest that kind of effort. You could ask Prof. Nash if he would be willing to do this, but if you don't have a compelling common use case, he probably is not going to either.