Haplotype assembly is the problem of reconstructing the combination of alleles on the maternally and paternally inherited chromosome copies and is key to our understanding of human population genetics and disease. Numerous statistical and molecular approaches have been developed to date to enable haplotype reconstruction. In this work, we focus on read-based phasing of individual genomes, which involves the assembly of the two haplotypes from whole-genome-sequencing read alignments and variant genotypes. Fragments that span more than one heterozygous variant provide molecular linkage evidence for alleles occurring on the same haplotype and can hence be leveraged for haplotype assembly; however, sequencing errors make this problem challenging. Existing techniques often employ an NP-hard combinatorial optimization formulation for this problem and rely on hand-engineered heuristics to find a solution. Here we propose a novel framework based on deep reinforcement learning, which integrates the representational power of deep learning with reinforcement learning, to automatically learn effective algorithms that can accurately partition read fragments into two haplotype sets given inputs from different sequencing platforms. Importantly, this approach does not require labeled training data, which allows us to use all the publicly-available datasets collected in large-scale sequencing repositories, such as the 1000 Genomes Project, as training data for our models. Given the complex combinatorial structure of genomic data, an important aspect of this work is the design and compilation of a representative training dataset to ensure model generalizability. Our initial preliminary results show that our approach can achieve state of the art phasing block lengths and lower error rates on short read inputs.