How the DeepSynergy.AI Planning Engine works
Planning knowledge, data and KPIs in – an optimised production plan out: how the engine works, from deep reinforcement learning to the digital twin.
The short answer: The DeepSynergy.AI Planning Engine is the optimisation core of DeepSynergy.AI Production Planning. It combines three inputs – your planners’ knowledge, your planning objectives as KPIs and your current planning data – and creates an optimised production plan from them. Creating and optimising a plan takes less than 30 minutes. Every plan is scored against your KPIs and checked for feasibility in a digital twin before it goes into production.
Three inputs, one plan
A good production plan does not come from data alone. The engine therefore works with three clearly separated inputs:
- A – Expert and process knowledge: rules, exceptions and priorities that often exist only in the heads of planners and supervisors. Which machine may process which order? When is an order ready to start? Which sequence saves set-up effort? How this knowledge is structured is explained in How your planners’ knowledge gets into the plan.
- B – Planning objectives and KPIs: what makes a plan good? On-time delivery, utilisation, set-up effort, inventory or lead time – and above all, how much weight each objective carries.
- C – Current planning data: orders, routings, machines, shifts, material and qualifications, typically from ERP and MES or from Excel files.
From these three inputs the engine creates a production plan, scores it against the KPIs and improves it step by step.
Why production planning is so hard to compute
Production planning is a combinatorial optimisation problem: there is a vast number of possible sequences and assignments of orders to machines, shifts and people, and that number grows explosively with every additional order. Such problems are known as NP-hard – at realistic sizes, it is impossible to try every option.
Conventional methods quickly reach their limits:
- Priority rules (such as “earliest due date first”) are fast and easy to follow, but ignore interactions such as set-up sequences or staff bottlenecks.
- Linear optimisation delivers exact results, but only for comparatively simple problems. Complex production can rarely be optimised exactly this way, and every process change means considerable effort.
- Problem-specific heuristics can deliver good solutions, but require deep mathematical expertise and costly adjustment whenever the process changes.
How the engine finds a good plan
For each planning problem, DeepSynergy.AI Production Planning combines suitable tools from a specialised technology stack. The key building blocks, explained without formulas:
- Deep reinforcement learning (DRL): an agent learns through trial and evaluation which planning decisions lead to good results. The “reward” is the plan’s score against your KPIs. This produces decision strategies that take many factors into account at once.
- Transformer networks: this network architecture recognises relationships between many elements at the same time – for example between orders competing for the same machine or the same material.
- Agent-based decision-making: instead of a rigid rule, a learning agent makes decisions in the plan step by step and evaluates their effect.
- Reducing problem size: methods such as autoencoders condense the planning situation to what matters, so optimisation stays fast even with many orders.
Verification in a digital twin
Because learning methods work with probabilities, every plan is then verified in a digital twin of the production. It checks whether the plan respects every constraint: machine capabilities, shifts, material availability and delivery dates. Only a verified plan is released.
Why not a language model?
Large language models (LLMs) such as ChatGPT are built for text, not for optimisation. They work probabilistically and cannot reliably guarantee that a plan meets every constraint or reaches a KPI optimum. A miscalculated delivery date or a wrong machine assignment quickly costs money in production. Production planning therefore needs specialised methods, domain knowledge and verification of every plan.
What this means for your planning
For planners, the engine is not a black box: the rules come from their own knowledge, the objectives are agreed together and every plan can be traced back through the KPIs. What the result looks like is described in The result: a production plan you can rely on. For an overview of the product, see the product page.
How much potential is there in your planning?
In the free potential check (45 minutes), we talk about your current planning and assess together whether and where DeepSynergy.AI Production Planning can help.
