In thefirst versions of the benchmark, a transmission con-trol protocol (TCP) interfacefor controllers was provided sothat controllers written in other languages than Java could be in-terfaced to the code. However, this TCP interface introduced aconsiderable communication overhead making non-Java agentsorders of magnitude slower, andhad occasional stability issues.For later releases, a new library calledAmiCowas developed forcommunication between the benchmark and agents developedin other language.TheAmiColibrary is applicable for inter-language processcommunication beyond the Mario AI benchmark. The purposeof the library is to provide an easy-to-use and as seamless aspossible bridge between foreign programming languages pre-serving high performance, comparable to the native languages’