Abstract: One example provides a recommendation system configured to receive, from a plurality of remote devices, user data corresponding to a plurality of users. The recommendation system is further configured to, for a user of interest, determine, using an optimizer comprising a machine learning model, a default likelihood of an outcome of interest being achieved based at least on a set of user actions of the user of interest that have occurred after a start point and to determine, using the optimizer, a first hypothetical likelihood of the outcome of interest being achieved based at least on the set of user actions and on a first possible user action being a next user action hypothetically performed. The recommendation system is further configured to, determine a next action to recommend to the user of interest, and output the next action to recommend.