Learning Local Forward Models on Unforgiving Games