Guessing Winning Policies in LTL Synthesis by Semantic Learning