Deep active inference agents using Monte-Carlo methods