Context-based Learning of Adaptive Vehicle Dispatching Strategies

The reinforcement learning framework has been used for finding and improving solutions to complex scheduling and resource allocation problems. However, in many real-world applications, state or action sets become very large due to the curse of dimensionality. This leads to approaches involving an approximation of the value function rather than a complete representation in a table. This contribution shows an application of reinforcement learning for a scenario in internal logistics involving multiple transportation vehicles operating in a common workspace. We show a solution for learning an adaptive vehicle dispatching strategy, which is suitable for a wide range of environmental conditions. This solution uses a context-based state representation that is scalable and allows a tabular representation of the value function even for large problem instances. The performance of the learned behaviour is evaluated by using a simulation model and is compared to common heuristic dispatching strategies.