Learning to Avoid Dominated Action Sequences in Planning for Black-Box Domains