Reinforcement Learning based Droplet Routing Technique in Hexagonal Digital Microfluidic Biochips using Dueling Network

Droplet route planning between the reconfigurable modules and modules to I/O ports is a very essential step in chip design and finding optimised routes is an NP-complete problem. Various routing algorithms like A* heuristic search-based algorithm, prioritised A* algorithm, network flow-based routing and parallel high-performance routing inspired by the graph colouring problem, etc. have been advanced to construe the droplet routing operation in the microfluidic realm. It is prominent that many of these previous studies are based on the prior knowledge of the route network. Here a reinforcement learning based Q learning technique is used to find out the optimal paths where the dueling architecture is used to form a dueling deep Q network throughout the chip array. The agent is the droplet, which learns through dueling deep Q network to avoid obstacles and outperforms the state-of-the-art on DMFB routing domain. In this proposed algorithm, for any obstacles in the droplet paths, the route compaction is concluded with the stalling and detouring methods. The simulation results exhibit better policy evaluation than previous approaches by avoiding unnecessary estimations of values under each action to the environment. The output value of the deep Q network is combined with the reward function and it is added to the reinforcement learning training in the guise of total reward.