Scheduling and operations / Reinforcement learning
Reinforcement Learning for Scheduling
Deep Q-Network approaches to scheduling decisions with consequences that unfold over time.
Development
Problem statement
Choose actions in a scheduling problem where each decision changes the options available next.
My contribution
Development of a reinforcement-learning approach to scheduling in collaboration with logistics teams.
Technical challenges
Domain-level considerations, without implementation-specific details.
- Representing a sequential decision problem in a tractable form.
- Balancing immediate objectives with longer-term consequences.
- Evaluating behavior against operational constraints.
Methodology
Use a Deep Q-Network to learn action values and inform sequential scheduling decisions.
- Express scheduling as a sequence of states, actions, and feedback.
- Evaluate candidate policies in the context of the scheduling objective.
Conceptual architecture
Abstract state
Action-value estimates
Candidate action
Evaluation and feedback
Feedback informs the next state and subsequent policy evaluation.
Technologies and methods
- Deep Q-Network
- Reinforcement Learning
- Scheduling