|Abstract:||Learning a grammar in Harmonic Serialism (Prince and Smolensky 1993/2004) requires identifying informative underlying forms and determining ranking information from them. Problematic from the learner's perspective is that not all input-output mappings are given, and some non-surface present forms and the forms they map to contain crucial ranking information about the target language. Tessier and Jesney (2014) clearly articulate the core of the matter: (1) which non-surface present underlying forms does one attempt to learn from, and (2) given such a form with an unknown output, what ranking information can be determined of the target language? This paper proposes answering these questions using two notions inherent to every HS system. First, HS systems organize their forms into a metric space, and this metric can be used to produce Local Candidate Maps. These maps provide an organizing tool for the learner to navigate the multitudinous search space of underlying forms. Second, informative ranking information can be obtained from an underlying form even when the learner does not know what it maps to. This is done by application of the join (Merchant 2008) to ranking information from possible maps. These two ideas are combined into a learning algorithm: Recursive Join Learning using Candidate Maps, that is shown to address many of the issues raised by Tessier and Jesney.