Distributed consensus is the core aspect of blockchain protocol security design. Recent protocols like IOTA have improved concurrency and scalability over Proof-of-work (PoW) with Bitcoin but have core design decisions that are inefficient for limited devices and do not take advantage of previous network experience to reduce calculations. This work proposes the first blockchain consensus protocol based on active machine-learning decisions, called Proof-of-history (PoH). PoH is setup as a distributed reinforcement-learning task for monitoring classification and training of blockchain transactions with an inner deep classifier. Early theoretical analysis and simulations show that PoH is robust to uncoordinated byzantine attacks through the use of a opinion-consolidation scheme for deep reinforcement-learning and weighted threshold voting scheme for decision-making. Future work will expand on these results to prove robustness and improve auditability with the deep networks.