In changing circumstances, success depends on making the right decision at the right time. Effective decision-making in dynamic and uncertain environments requires integrating future predictions into the decision-making process and acting in accordance with constraints. In daily business life, we make dozens of small decisions such as which task to prioritize, how to allocate resources, how much to produce, or how to manage energy. However, when machines, budgets, capacity limits, and changing conditions come into play, the "right decision" suddenly becomes much more complex: on one hand, there are tasks that need to be completed and cost pressures, and on the other hand, there is limited capacity, variable demand, and rapidly changing conditions. Moreover, the issue is not just about managing today; it is necessary to prepare for possible scenarios and foresee risks in order to continue without interruption when conditions change tomorrow. At this point, the project aims to develop a framework based on Deep Reinforcement Learning (DRL), an Artificial Intelligence method, that combines these two needs, works in accordance with predictions and constraints, and produces smarter, more agile, and implementable decisions. Within this framework, the project;
- By incorporating temporal information (e.g., patterns in signals such as demand, temperature, and cost) into policy learning, we aim to develop adaptive decision-making policies under uncertainty using SDP algorithms.
- Lagrangian relaxation allows for the flexible inclusion of operational constraints such as capacity and budget in the decision-making process.
- To support the collaborative learning and coordination of multiple units (e.g., machines/lines/boilers) through a multi-agent learning approach,
- Graph Neural Networks aim to contextually model resource-task (e.g., machine-product) relationships and integrate them into the learning process.
Making decisions under uncertainty requires much more than simply analyzing the current situation. Modern production systems are dynamic structures where variable demand, fluctuating costs, environmental impacts, and operational constraints all have an effect simultaneously. Developing an effective decision-making mechanism in such environments necessitates an approach that can evaluate different information sources together and adapt by learning over time.
This project aims to develop an integrated framework that integrates multiple data sources and supports the decision-making process through artificial intelligence-based learning mechanisms. The figure summarizes the overall flow of this integration. The system takes into account demand behaviors and temporal patterns observed in the past.
This information is incorporated into the decision-making process through in-depth time series analysis. When external conditions such as cost and price change suddenly, a decision that was correct yesterday may not work today; therefore, environmental information is directly brought into the decision-making process. Constraints such as capacity, energy, and budget are not constraints that are controlled later, but are part of the framework that shapes the decision from the outset; thus, the policies produced are not only effective but also implementable.
All these information flows converge in a deep neural network-based learning core, an AI method, shaping how the system makes decisions. The mechanism learns better strategies over time by experiencing the consequences of decisions made under different conditions; thus, decision-making ceases to be a process based on fixed rules and transforms into a policy structure that can be updated under uncertainty. The resulting policy aims not only to learn decisions that "look good" in the short term, but also to develop decision-making habits that are feasible under constraints and can withstand changing conditions.
To translate these decision-making habits into operational reality, the learning process is conducted under numerous scenarios on a simulation representing the operational environment. As the decisions generated by the policy are executed in the simulation, the system observes performance responses and constraint pressures; these observations are reintroduced into the learning cycle as feedback, and the policy is updated. Thus, instead of generating plans from scratch with every change in conditions, the approach gains an adaptable structure that improves itself through scenario-based experience accumulation. The same learning logic is extended to encompass the simultaneous and coordinated progress of multiple units in multi-resource systems; coordination is therefore treated as a natural component of the process. As a result, the framework provides a general methodological foundation that produces applicable and adaptive decisions under uncertainty; it aims to develop policies that are adaptable to changing conditions for problems such as production planning, capacity/energy management, and resource allocation.
