Federated Discrete Reinforcement Learning for Automatic Guided Vehicle Control

Under the federated learning paradigm, the agents learn in parallel and combine their knowledge to build a global knowledge model. This new machine learning strategy increases privacy and reduces communication costs, some benefits that can be very useful for industry applications deployed in the edge. Automatic Guided Vehicles (AGVs) can take advantage of this approach since they can be considered intelligent agents, operate in fleets, and are normally managed by a central system that can run in the edge and handles the knowledge of each of them to obtain a global emerging behavioral model. Furthermore, this idea can be combined with the concept of reinforcement learning (RL). This way, the AGVs can interact with the system to learn according to the policy implemented by the RL algorithm in order to follow specified routes, and send their findings to the main system. The centralized system collects this information in a group policy to turn it over to the AGVs. In this work, a novel Federated Discrete Reinforcement Learning (FDRL) approach is implemented to control the trajectories of a fleet of AGVs. Each industrial AGV runs the modules that correspond to an RL system: a state estimator, a rewards calculator, an action selector, and a policy update algorithm. AGVs share their policy variation with the federated server, which combines them into a group policy with a learning aggregation function. To validate the proposal, simulation results of the FDRL control for five hybrid tricycle-differential AGVs and four different trajectories (ellipse, lemniscate, octagon, and a closed 16-polyline) have been obtained and compared with a Proportional Integral Derivative (PID) controller optimized with genetic algorithms. The intelligent control approach shows an average improvement of 78% in mean absolute error, 75% in root mean square error, and 73% in terms of standard deviation. It has been shown that this approach also accelerates the learning up to a 50 % depending on the trajectory, with an average of 36% speed up while allowing precise tracking. The suggested federated-learning based technique outperforms an optimized fuzzy logic controller (FLC) for all of the measured trajectories as well. In addition, different learning aggregation functions have been proposed and evaluated. The influence of the number of vehicles (from 2 to 10) on the path following performance and on network transmission has been analyzed too.