This paper proposes a learning method to select the most appropriate abstract plans during hierarchical planning in the context of multi-agent systems (MAS). In hierarchical planning, a plan is first created at the most abstract level, and is then refined to a more concrete plan, level by level. Thus, selecting an appropriate plan at the abstract level is very important because the selected plan restricts the scope of lower concrete-level plans. This restriction can enable agents to create plans efficiently, but if all the plans under the selected plan contain serious and difficult-to-resolve conflicts with other agents' plans, the resulting plan does not work well or is of low quality. We propose a method in which, from the conflict pattern among agents' plans, an agent learns which abstract plans will cause conflicts with less probability or which conflicts are easy to resolve, thus inducing probabilistically higher-utility concrete plans after conflict resolution. We also show some experimental results to evaluate our method, with the results suggesting structures of resources where tasks are executed.