This paper considers a mobile multi-agent system (e.g., a crowdsensing or crowdsourcing application) in which agents are characterized by their discrete-valued Level of Ability (LoA) at doing some sensing or data processing task. Agents are not aware of their LoA and are willing to estimate it without the help of a central classification authority, in order to determine whether their contribution will improve or degrade the performance of the global network. Using their estimated LoA, agents may then voluntarily restrain themselves to participate to the activity of the multi-agent system, without being banned by some central control authority. For that purpose, agents, when they meet, perform pairwise comparison tests (PCT) able to determine which is the best agent of a pair of competing agents. Two maximum \emph{a posteriori} (MAP) estimators of the LoA of each agent are proposed using the results of several PCTs. These MAP estimators are then employed to determine, for a given agent, the LoA of the next opponent that minimizes its probability of LoA estimation error. Simulations results show that the proposed optimal opponent selection approach provides better results than simply choosing opponents at random, or always choosing the opponents with the best LoA.