Determining the number of contributors and deconvolving mixed Short Tandem Repeat (STR) profiles represent critical challenges in forensic DNA analysis. Existing methods struggle to simultaneously achieve accurate contributor estimation and stable profile deconvolution when processing complex mixtures with an unknown number of contributors. To address these challenges, this paper proposes a method for contributor identification and intelligent deconvolution of mixed STR profiles based on Hierarchical Multi-Agent Reinforcement Learning (H-MARL). The proposed method constructs a two-level progressive decision-making framework, decoupling the complex deconvolution task into two synergistic levels: high-level global inference and low-level local deconvolution. In the high-level inference network, the model captures the global representation of the STR profile and employs a dual-branch structure to synchronously output a discrete prediction of the number of contributors (K) and the prior distribution of mixture proportions. In the low-level collaborative network, the system dynamically activates a corresponding number of agents based on the K value issued by the high-level network, mapping each agent to a potential DNA contributor. Based on the allelic sequence features of the loci, the agents employ the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm to make collaborative decisions. During the training phase, the agents undergo global optimization, while in the inference phase, they execute independently. Through a message-passing mechanism, the agents efficiently complete the exclusive allocation of alleles at each locus and fine-tune mixture proportions. Experimental results demonstrate that the proposed method achieves high accuracy in contributor identification and stability in profile deconvolution in scenarios with an unknown number of contributors. It maintains robust performance even under complex noise conditions, providing a novel research avenue for the intelligent interpretation of complex mixed STR profiles with unknown contributor counts.

Authors

• Liang Zeng (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Shanping Yu (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Zhehua Mao (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)