Are large windows efficient? Evolution of learning rules in a bargaining model