Manufacturing and FMCG enterprises invest millions in enterprise resource planning (ERP) systems to better manage operations. Yet on the factory floor, the agility keeping lines running often relies on reactive firefighting and workarounds.
This operational contradiction is the Agility Paradox. Core ERP platforms assume a static world based on fixed averages. Real production environments, however, are governed by dynamic physical constraints: sequence-dependent changeovers, fluctuating labour availability, shared engineering resources, and sudden demand shifts. When faced with disruption, rigid schedules break. Factory managers face a choice between sticking to plans that idle machinery, or best-guessing and spreading chaos down the line.
How Digital Twins Unlock True OEE and De-Risks Operations
Bridging the gap between macro-level enterprise planning and shop-floor volatility requires Simulation-Based Optimisation and Digital Twins. Rather than a static 3D model or retrospective dashboard, a digital twin provides a living computational replica of your facility’s physical constraints and operational logic.
Integrating simulation delivers immediate strategic and tactical advantages:
Predictive Bottleneck Prevention: Evaluate alternative shift patterns, line allocations, and batch sequencing in a virtual sandbox to resolve bottlenecks ahead of time.
CAPEX De-Risking: Test new machinery, packing stations, and line layouts virtually to validate throughput and present boardrooms with ROI proof before issuing purchase orders.
Upstream Complexity Management: Quantify the operational and financial impact of complex orders and sequencing before scheduling production.
Rapid Commissioning: Accelerate ramp-up timelines and minimise downtime during major asset installations.
Validated in High-Stakes Production
Simulation-based decision intelligence is already proven across complex UK and European supply chains. Recipe-box pioneer Gousto used Decision Lab’s simulation engine as a digital sandbox to optimise network throughput and scale rapidly without sacrificing service levels. In highly regulated life sciences, global leaders like GSK and AstraZeneca deploy advanced simulation to de-risk multi-stage batch processes and maximise asset utilisation under zero-fault conditions.
Dive Deeper in the Complete Whitepaper
Elevate planning and eliminate the drag of operational guesswork. Learn how to transform plant complexity into predictive agility. Dive deeper, with examples and case studies, in our Agility Paradox paper.
AnyLogic is a powerful simulation tool that can be used to simulate a wide range of systems and processes. It allows developers to build simulations that take in a user defined set of inputs and provide a robust set of outputs. Therefore, it can be a useful tool to aid in optimisation problems by taking in a set of inputs and quickly generating outputs that can then be evaluated to determine how well these inputs performed. Over multiple runs of the simulation with varying inputs you can determine which inputs provide the optimal results.
AnyLogic provides two main types of experiment that can be used for optimisation. Firstly, there is the Optimisation Experiment which is built on top of the OptQuest Simulation Engine and is quick and easy to set up within AnyLogic. Secondly, there is the Custom Experiment which is entirely specified by the user and gives them full control and flexibility over how the simulation runs. However, to perform optimisation with Custom Experiment it requires the user to set up an external Java programme which calls and runs it. Another area that is useful to consider is how accessing python optimisation scripts can enhance an AnyLogic simulation and allow it to make better decisions or perform more complex processes. The best method for this is using the Java Pypeline library which provides AnyLogic with a way to make calls to external python scripts.
Optimisation Experiment
The Optimisation Experiment is a built-in feature within AnyLogic which is designed for running multiple scenarios to try and minimise/maximise the ‘objective function’. The objective function is the value that we want to optimise such as minimising lead time or maximising revenue. This could be a single output or multiple outputs that are resolved into a single aggregated value. The user can specify a set of inputs that are either fixed or are varied between runs of the model so that the best values can be determined. By varying inputs, such as the number of available resources, the model can determine the optimal set of inputs. There is also built-in functionality for performing runs of the model in parallel and evaluating multiple runs of the same inputs to account for randomness within the model such as varying levels of demand for resources at different times. These are good out-of-the box features that are easy to use and can provide a lot of value for simple optimisation problems.
What are the Limitations of Native AnyLogic Optimisation?
The out-of-the-box features provide immediate value for straightforward scenarios, the native setup falls short for highly complex engineering or logistics problems.
Algorithmic Constraints: The native experiment types are strictly limited to minimising or maximising a specific value. Advanced heuristic search techniques, such as the Hill Climbing method, are not natively supported.
Parameter Structure Restrictions: The optimiser cannot natively vary complex inputs like arrays or lists.
Consequently, varying complex data structures must still be managed manually between runs. Developers may find it necessary to implement dummy inputs simply to ensure the optimiser completes, subsequently collecting performance statistics manually rather than relying on the engine’s built-in outputs.
In theory, a developer can write custom code within the experiment to execute algorithms like Hill Climbing. However, this approach essentially bypasses the actual OptQuest engine, utilising the platform’s multi-run framework as a workaround for something that could be achieved just as effectively using a standard Monte Carlo experiment.
Figure 1: Example of the default Optimisation Experiment UI
Custom Experiments: The Cloud-Native Paradigm
While built-in experiments offer out-of-the-box simplicity, the Custom Experiment framework unlocks total flexibility, granting full programmatic control over simulation logic, execution parameters, and model run-times.
Historically, running a custom experiment meant orchestrating a cumbersome, local Java programme to manually trigger consecutive simulation iterations. Today, this restrictive approach has been entirely replaced by a cloud-native paradigm. Rather than binding computation to a developer’s local machine, modern architectures decouple the simulation model from the optimisation logic.
Scaling with the AnyLogic Cloud RESTful API
The contemporary standard for advanced orchestration relies on the AnyLogic Cloud RESTful API, which features native client libraries for Python, Java, and JavaScript. This multi-language web interface fundamentally transforms how custom optimisation loops are executed:
Serverless-Style Execution: Developers can completely bypass local environmental constraints. Simulation models are uploaded to a cloud or private server instance, where individual runs are triggered headlessly and asynchronously on-demand.
Distributed Parallelisation: Highly intensive, multi-run optimisation loops are offloaded to high-performance infrastructure. The cloud platform splits the workload horizontally across distributed cloud nodes and cores, executing thousands of scenarios simultaneously and drastically shortening time-to-insight.
Language-Agnostic Heuristics: By utilising the Cloud API, engineers are no longer forced to write complex search heuristics exclusively in Java. A data scientist can now design sophisticated search algorithms (such as the Hill Climbing method, genetic algorithms, or deep reinforcement learning policies) directly within a Python script, leveraging powerful open-source libraries like Google OR-Tools, SciPy, or PyTorch.
For example, a Python orchestration script can execute a supply chain simulation via a clean REST call, instantly analyse the returned outputs to pinpoint a localised bottleneck (such as a depleted inventory layer or an asset constraint), update the input vectors, and immediately fire a new flight of parallelised runs across the cloud cluster.
This modern cloud workflow allows businesses to use their AnyLogic digital twins as highly scalable, serverless objective function evaluators for cutting-edge external solvers.
Figure 2: Example of the Custom Experiment Set UpFigure 3: Diagram of workflow for running Custom Experiment from external optimiser programme.
Hybrid Decision Intelligence: Choosing Your Architectural Pattern
When moving beyond native probing limits, enterprise models require a hybrid approach that pairs discrete-event simulation with sophisticated mathematical solvers or analytical models. Depending on your tech stack, latent memory requirements, and data science infrastructure, two primary patterns emerge.
Pattern A: The Analytics & ML Bridge (Figure 4)
Figure 4: Diagram showing workflow of using an optimiser to supplement an AnyLogic model.
While using native Java remains the foundation for internal scripting, enhancing an AnyLogic simulation with Python’s vast ecosystem of data science, optimisation, and machine learning libraries is now standard practice. However, Python integration is no longer a one-size-fits-all workaround. Modern architectures deploy a structured, dual-pathway approach depending entirely on which environment dictates the execution flow.
Pathway 1: Pypeline (v1.9.6) – The Python-inside-Java Architecture
The Pypeline custom library is the premier tool for workflows where the AnyLogic model acts as the parent process. In this configuration, the simulation runs normally in its Java environment but reaches out to a local Python installation to execute discrete, targeted scripts on demand.
Best For: Scenarios where the simulation requires real-time assistance from a Python library to make an internal decision (such as querying a pre-trained scikit-learn model to predict an asset failure rate mid-run or calling Google OR-Tools to solve a complex Travelling Salesman problem for fleet routing).
The Jsonifier Advantage: Modern iterations of Pypeline are deeply integrated with the Jsonifier serialisation framework. This eliminates the historical headache of manually converting Java data structures into standard text strings. Jsonifier allows entire agent populations, multidimensional tables, and complex database tuples to be seamlessly serialised into machine-readable JSON strings and deserialised back into native Java objects with minimal computational overhead.
Figure 5: Default settings for the Pypeline communicator object. Allows you to select which version of python.
Pathway 2: Alpyne (v1.2.1 Beta) – The Java-inside-Python Architecture
When the orchestration flow needs to be completely reversed, the Alpyne (v1.2.1 Beta) library provides a powerful alternative. Under this paradigm, the master Python script acts as the host environment, treating the AnyLogic simulation as an importable, interactive object.
Best For: Advanced AI training, complex heuristic loops, and intensive Reinforcement Learning (RL) pipelines.
How it Works: Developers export their model from AnyLogic using the native Reinforcement Learning experiment framework. The Alpyne package then initialises a standalone AnyLogicSim connection inside Python. The Python script can interactively step the simulation forward, freeze execution at pre-defined decision points, inspect the exact internal observation state, inject an algorithmic action, and collect rewards.
AI Testbed Compatibility: Because Python serves as the control room, engineers can natively wrap the simulation inside standard machine learning frameworks (like OpenAI Gym/Gymnasium) to train neural networks using deep RL frameworks like PyTorch, TensorFlow, or Stable-Baselines3.
Pattern B: The High-Performance Embedded Engine (Figure 6)
Embedding the Gurobi Optimiser for Mathematical Rigour
For large-scale enterprise problems (such as global supply chain networks, complex asset-scheduling, or multi-echelon inventory routing) standard heuristics and search algorithms often struggle to converge on a true mathematical optimum. In these scenarios, the highly robust architecture option involves embedding a high-performance mathematical solver like Gurobi directly into the simulation environment.
Because AnyLogic is natively built on Java, developers can easily add the gurobi.jar dependency directly into the simulation project’s build path. This setup unlocks a powerful, two-way hybrid framework:
Balancing Stochasticity with Deterministic Rigour: AnyLogic excels at capturing the messy, unpredictable realities of a system (stochastic delays, machine breakdowns, fluctuating demand). Meanwhile, Gurobi excels at solving dense, deterministic mathematical models with millions of variables and constraints.
Dynamic, Mid-Simulation Re-optimisation: Rather than just using an optimiser to find inputs before a model runs, an AnyLogic model can call the Gurobi Java API during a run. For example, every simulated 24 hours, the model can halt, pass its current state (such as unexpected delivery disruptions or machine outages) to Gurobi as a matrix, and allow Gurobi to instantly calculate a mathematically optimal production schedule or fleet routing blueprint for the next day. The simulation then resumes using that flawless plan.
By combining AnyLogic’s granular, event-driven simulation with Gurobi’s industry-leading prescriptive analytics engine, engineers can construct highly resilient digital twins that not only accurately predict system bottlenecks but actively solve them with absolute mathematical certainty.
Micro-latency (Optimised for macro-step decisions)
Near-Zero Latency (Optimised for high-frequency loops)
Key Software Stack
AnyLogic + Python (scipy, scikit-learn, OR-Tools)
AnyLogic + Gurobi / CPLEX Native Java API
Final Thoughts: Choosing Your Optimisation Architecture
Selecting the right optimisation pathway in AnyLogic depends entirely on the scale of your problem and the complexity of the data structures involved. There is no single “correct” approach, but rather a spectrum of engineering choices tailored to enterprise requirements:
For Straightforward Parameters: If your objective is simply to vary a handful of discrete or continuous scalar variables to find a clear minimum or maximum, the native Optimisation Experiment offers a highly capable, out-of-the-box framework driven by OptQuest that requires minimal development overhead.
For High-Performance Cloud Scaling: When optimisation loops involve multi-dimensional arrays or require advanced custom search heuristics, pivoting to a Custom Experiment executed via the AnyLogic Cloud RESTful API is the modern benchmark. This allows you to offload computational strain serverless-style across distributed cloud nodes.
For Ecosystem Integration: If your solution relies on data science or machine learning libraries, deploying a structured, dual-pathway approach via Pypeline (v1.9.6) or Alpyne (v1.2.1 Beta) ensures a robust data exchange between Java and Python, depending on whether the simulation or an external script controls the parent execution flow.
For Absolute Mathematical Rigour: For highly dense, combinatorial problems where heuristics fail to guarantee a true global optimum, embedding a premium solver like Gurobi directly into the simulation code path unlocks powerful math-heuristic capabilities. This empowers digital twins to balance unpredictable real-world stochasticity with precise, mid-run prescriptive analytics.
By understanding these structural boundaries, engineering teams can stop treating simulation models as passive visual tools and start utilising them as scalable, highly intelligent decision-support engines.
Transform Your Operational Complexities into Provable Success
At Decision Lab, we specialise in bridging the gap between intricate system simulations and industry-leading mathematical optimisation. Whether you are aiming to de-risk global supply chains, design high-efficiency logistics networks, or train cutting-edge reinforcement learning agents, our decision intelligence experts can help you architect the perfect technical framework.
Decision Intelligence moves an organisation beyond the fixed-plan trap toward proactive, automated resilience in route optimisation. By evaluating the strategic trade-offs between explainable Stochastic Optimisation and scalable Reinforcement Learning, we proved that move-level agility is the key to maintaining flow in a high-uncertainty environment.
Key Takeaways:
Beyond Rigid Scheduling: Shifting from historical templates to dynamic, operational-time decision-making to maximise capacity utilisation.
The Technical Showdown: Comparing the audit trails of Stochastic Programming against the autonomous adaptability of Reinforcement Learning.
Predictive Simulation: Utilising a road-based “digital sandbox” to test courses of action and mitigate risks before committing resources.
Tangible ROI: Delivering financial returns by improving linehaul utilisation and significantly reducing the need for costly ad-hoc transport.
The Challenge: The Friction of Fixed Planning
In the high-stakes corridors of European logistics, fixed plans are often the first casualty of reality. For an Operations Director managing a distribution network across the UK and EU, the daily friction is visceral. You are constantly forced to ask: “Should I delay this trailer, so it leaves full, or stick to the schedule? Do I need to commission an expensive ad-hoc truck to cover this surge, or will the bottleneck clear itself?”
When package volumes fluctuate unpredictably at major hubs, static schedules become more than just an inconvenience—they become a drain on margins and a threat to service levels. At Decision Lab, we operate under a foundational truth: the success of an organisation is nothing but the sum of all its decisions. To help global leaders move beyond the fixed-plans trap, we conducted a six-week Proof of Concept (PoC) with FedEx. This project tackled real-world complexity head-on, proving that Decision Intelligence is the key to transitioning from reactive firefighting to proactive, automated resilience.
Static Schedules are the Enemy of Efficiency
The core challenge identified within FedEx’s European network was the inherent limitation of pre-defined linehaul schedules. These schedules were designed for averages, whereas logistics are often defined by exceptions. When incoming and outgoing package volumes at European hubs diverged from the forecast, rigid plans could not adapt.
A dynamic approach, powered by operational-time decision-making, is the only way to maintain flow in a volatile environment. By rerouting assets and scheduling departures based on real-time parcel traffic rather than historical templates, an organisation can achieve step-changing improvements in capacity utilisation.
Our expertise in AI, ML, simulation, and mathematical optimisation helps organisations cut through complexities in strategic, tactical and operational processes.
The Solution: Bridging the Gap with Decision Intelligence
The choice between technical approaches is rarely straightforward; in this case it was a strategic balancing act between Explainability and Scalability. During our PoC, we evaluated two competing methodologies: explainable Stochastic Programming and scalable Reinforcement Learning (RL).
Feature
Stochastic Programming
Reinforcement Learning (RL)
Primary Strength
Fast solving speed; mathematically explainable and provable.
Reacts to high uncertainty using World Models and Graph Neural Networks.
Logic Basis
Locates the best strategy to optimise expected outcomes over uncertainty.
Uses a dynamics model to predict the optimum next action.
Adaptability
Multi-objective handling: Uniquely suited for balancing cost vs. customer service levels.
Observation-size invariant: Handles environments with variable data lengths and network nodes.
Strategic Risk
Consulting intensive: Very sensitive to human-built heuristics, which are expensive and time-consuming to develop.
Compute intensive: Requires significant hardware resources for training the World Model.
While Stochastic Programming offers a clear audit trail for every decision, RL provides the adaptability required for massive, interconnected networks. The right choice depends on whether your organisation prioritises a provably optimal solution or a highly performant best-effort that can autonomously learn the shifting dynamics of global markets.
Don’t Just Predict—Simulate the Impact
One of the most powerful tools developed for FedEx was a road-based, hub-to-hub package movement simulator. This provides a digital sandbox where controllers can explore alternative COAs (Courses of Action) before committing resources.
Our completely data-driven deployment method allows us to build these simulations without the months of manual coding traditionally required. By accessing relevant operational and transport data directly, we can simulate supply chain environments to predict the ripple effects of a delay or reroute.
This tool predicts the impact of different actions, helping to mitigate risks and optimise routes.
Data Maturity is the Ultimate Competitive Moat
For large-scale firms with a £200M+ turnover, the transactional backbone—usually an ERP or MRP system like SAP or Oracle—is necessary but insufficient. To achieve true antifragility, you must layer Decision Intelligence over these systems.
Antifragility is the ability to not just survive volatility, but to actually improve because of it. By utilising a World Model within an RL framework, the system treats every fluctuation in package volume as a learning opportunity, refining its dynamics model to better anticipate future shocks. This requires three layers of data maturity:
Strategic Level: Long-term high-level routes, fleet capacity, and cost-per-mile data.
Operational Level: Real-time visibility into items currently loaded or waiting at the depot.
Historical Level: Deep archives of how volumes fluctuated in similar time slots in the past.
The Result: Antifragility and Bottom-Line Returns
In the C-suite, the value of AI is measured by the bottom line. The FedEx project was not an academic exercise; it was focused on delivering the financial returns demanded by an industry with tight margins. The PoC demonstrated that an autonomous planning agent directly impacts:
Improved Linehaul Utilisation: Driving higher Overall Equipment Effectiveness (OEE) across the fleet.
Reduced Rescheduling: Eliminating the administrative friction and cost of mid-stream plan changes.
Minimised Ad-hoc Linehauls: Directly de-risking Operational Expenditure (OPEX) and informing more accurate Capital Expenditure (CAPEX) by reducing the need for emergency transport.
Conclusion: Toward the Global Digital Twin
The ultimate evolution of this journey is an advanced road-based package movement digital twin. By connecting multiple hubs in real-time, organisations can create a living model of their entire network that learns, adapts, and optimises itself.
What is the sum of your organisation’s decisions? How many of your current logistics choices are being left to a fixed plan that no longer fits your reality? In a world of increasing volatility, the goal is no longer just to have a plan—it is to have a system that provides decision clarity and reliable value.
Transform your logistics operations today. Reach out directly via out contact page, or connect with us on LinkedIn to start a conversation about de-risking your future.
As we navigate the technological landscape of 2026, Generative AI has undoubtedly transformed the way we interact with information. Chatbots and Large Language Models (LLMs) have proliferated across enterprise software, streamlining communication and automating basic workflows. However, for operations and supply chain leaders in complex, capital-intensive industries like FMCG, Automotive, and Retail manufacturing, a stark reality is emerging: LLMs are not a silver bullet.
While language models excel at processing text, they cannot single-handedly optimise a global supply chain network, nor can they provide the quantitative assurance needed to de-risk a £50m factory expansion. When dealing with physical realities, extreme market volatility, and fragmented legacy systems, text prediction is insufficient.
By 2026, 75% of Global 500 companies will apply decision intelligence practices
Gartner
The definitive competitive edge in 2026 belongs to those looking beyond Generative AI toward Decision AI. It belongs to organisations thoughtfully advancing their tech stack and building on established capabilities to embrace the architecture of a Composable Decision Intelligence Platform (DIP).
The Industrial Reality: High Stakes and High Volatility
Traditional planning systems struggle to account for agile and accelerated business and the consequent hazards. Today’s supply chain and operations leaders are caught in a crossfire of overlapping challenges, two of the most critical being:
Demand & Supply Volatility: SKU proliferation, shifting consumer behaviours, and frequent supply chain disruptions are breaking static planning models. The inability of legacy systems to cope with this extreme volatility inevitably results in poor service levels, excess inventory, and spiralling costs.
High-Stakes CAPEX Uncertainty: Securing funding for major capital investments—whether a new automated line, a facility expansion, or rationalising a post-merger manufacturing network—requires robust, data-driven justification. Without quantitative assurance, it is incredibly difficult to de-risk these investments and guarantee ROI.
Solving these multi-dimensional problems requires more than just analysing past data; it requires a platform capable of simulating the future and discovering the optimal path forward.
The Path to Implementation: A Composable Architecture
At Decision Lab, we deliver Decision Intelligence to help leaders master this uncertainty. We achieve this not through a rigid, black-box AI model, but by building a Composable Decision Intelligence Platform based on responsible AI TRiSM principles.
Composability is the principle that enables businesses to be agile. Rather than relying on a single vendor’s inflexible suite, a composable DIP orchestrates best-in-class, modular capabilities that ingest data from fragmented ERP, MES, and WMS systems. This creates a unified, dynamic view—an AI Simulation Twin.
The Strategic Advantage of the AI Simulation Twin
Instead of waiting years for a fully instrumented, hardware-dependent Digital Twin, leading organisations are accelerating their time-to-value by deploying an AI Simulation Twin.
Traditional Digital Twin programmes often stall in pilot purgatory due to immense IoT integration challenges, prohibitive hardware costs, and fragmented legacy data pipelines. A Simulation Twin, while still ingesting real data, fundamentally bypasses these immediate infrastructure hurdles. It delivers the core predictive and prescriptive advantages now—providing a high-fidelity virtual environment to solve urgent CAPEX and operational bottlenecks—while your physical IoT maturity can be developed as a separate, parallel track. This decoupling ensures you realise ROI in months, rather than years, before moving into the four pillars of the platform:
1. The Cognitive Engine: Autonomous AI Agents
Agentic AI serves as the reasoning layer of the platform. These agents can interpret complex scenarios, model market volatility, and process multi-tiered supply chain dynamics, translating raw data into actionable context.
2. The Virtual Sandbox: Simulation
To understand a complex physical network, you must be able to interrogate it and test it. Practically, that means replicating it. We use simulation to build a high-fidelity digital twin environment, employing appropriate technologies, such as AnyLogic’s multi-method capabilities. A simulation maps constraints, machines, and distribution nodes, providing the holistic view necessary to test what-if scenarios safely. It answers critical CAPEX questions before money is spent.
3. The Mathematical Engine: Optimisation
Where simulation shows you what could happen, optimisation dictates what should happen. For us, that means employing mathematical optimisation, such as Gurobi’s world-class mathematical solver, to cut through millions of potential permutations. It discovers the mathematically perfect production schedules and inventory policies—maximising throughput and service levels while minimising duplicated costs. The key is being timely—it is no good getting the answer after it was needed. Gurobi’s speed is key here (Gurobi white paper on solver speed).
4. The Continuous Learning Loop: Reinforcement Learning
This is where the platform moves from a passive analytical tool to an active operational asset. By applying Reinforcement Learning, specifically leveraging AgileRL, a platform can learn from real-time feedback. It continually experiments within the simulation, discovering new strategies to navigate supply shocks or demand spikes as they happen.
Engineering the Antifragile Supply Chain
The ultimate goal of implementing a Composable Decision Intelligence Platform is to shift operations from a state of fragility to one of Antifragility.
A robust system merely survives a shock. An antifragile operational system improves when exposed to volatility. When a sudden supply chain disruption occurs, the reinforcement learning algorithms immediately assess the new reality within the simulation, trigger the optimisation engine to recalculate the best path, and deploy autonomous agents to orchestrate a self-adapting response. Relying on singular AI models or monolithic ERPs to solve complex physical problems is being consigned to the past.
For leaders navigating constant disruption, true agility requires an adaptable, composable ecosystem. By implementing a Decision Intelligence Platform, you gain the foresight not just to predict the future, but to engineer your position—a compelling competitive advantage now and for the future.
In the world of complex decision-making, organisations often rely on two distinct tools. On one hand, there is Simulation (‘What happens if…?’), allowing us to model uncertainty and test scenarios. On the other, there is Optimisation (‘What is the best choice?’), allowing us to find the ideal solution within constraints.
Separately, they are powerful. But when integrated, they unlock a new level of capability—moving from simple decision support to intelligent, autonomous systems.
At Decision Lab, we don’t believe there is one single best method for this integration. The ideal approach depends entirely on the business problem at hand. Below, we explore the three primary patterns we use to drive value for clients like Migros, FedEx, and Nestlé, and how these models contribute to building truly antifragileorganisations.
Three Patterns of Integration
We generally view the integration of simulation and optimisation across a spectrum, moving from tactical support to full autonomy.
1. Optimisation within Simulation (Complex Decision Support)
In this pattern, the simulation runs a large-scale system, such as a warehouse. When a complex, real-time decision is required, the simulation pauses to call a dedicated optimisation algorithm.
How it works: The algorithm solves the specific sub-problem, and the simulation continues, testing how that “optimal” decision performs under real-world uncertainty (like worker delays).
Case Study: For Migros, we utilised this method. Their warehouse simulation calls an optimisation algorithm to determine the most efficient trolley-picking route every time a new order arrives. This allows us to test the routing logic’s real-world impact on the system’s total throughput. Case, video.
Here, the roles are reversed. An external optimisation wrapper searches for the best strategic solution, such as a factory layout or supply chain network.
How it works: For every solution the optimiser proposes, it uses the simulation as a high-fidelity “evaluation function” to test performance against stochastic conditions.
Case Study: For DataForm Lab, an optimisation model proposed various wind farm layouts. Our simulation then tested each layout against uncertain wind and wave conditions to calculate true energy output. The optimiser used this feedback to find the next, better solution.
3. Simulation trains Optimisation (The Autonomous Future)
This is where we enter the realm of the Digital Twin and Reinforcement Learning (RL). The simulation acts as a high-speed, risk-free training environment.
How it works: A machine learning agent interacts with the simulation millions of times, learning an ‘optimal policy’ for making autonomous decisions.
Case Study: For FedEx, we built a simulator for their linehaul operations. An AI agent was trained inside this simulator to learn the optimal policy on when to “cancel, delay, or add” linehauls based on uncertain package volumes, dramatically improving efficiency.
Building Antifragility: Beyond Resilience
Why go through the effort of building these combined models? It isn’t just about efficiency; it is about survival and growth.
A key philosophy at Decision Lab is Antifragility. While resilient systems merely withstand shock, antifragile systems improve because of it. By integrating simulation and optimisation, we create a “Digital Twin” that acts as a long-term asset.
We use these models to perform rigorous Sensitivity Analysis—identifying which inputs drive outcomes—and to stress-test operations against millions of potential scenarios. This allows organisations to design supply chains and operations that are prepared not just for the average day, but for uncertainty and volatility.
Every powerful methodology has trade-offs. We mitigate these through a rigorous Verification & Validation (V&V) process.
Computational Cost: Evaluating thousands of simulation runs is intensive. We use sensitivity analysis to identify key variables early, reducing the search space and focusing computational effort where it matters.
Data Dependency: “Garbage in, garbage out” applies doubly here. We don’t just use average values; we statistically analyse historical data to find correct probability distributions (e.g., ensuring orders follow specific peak-and-trough patterns, not just flat averages).
The ‘Human-in-the-Loop’
Ultimately, we build these models with you, not just for you.
Our methodology relies on a Human-in-the-Loop (HITL) framework. Whether we are helping Gousto compress development time for routing logic from months to days, or helping Nestle optimise warehouse slotting, the goal is the same: to present insights that empower expert judgment, not replace it.
Ready to build your Digital Twin?
To ensure a successful project, we look for four preliminary conditions:
A clear, quantifiable business problem.
Access to operational and historical data.
Dedicated engagement from your Subject Matter Experts (SMEs).
Clearly defined system boundaries.
If you are ready to move beyond simple guesswork and start engineering an antifragile operation, Contact Decision Lab today.
Manage Consent
To provide the best experiences, we use technologies like cookies to store and/or access device information. Consenting to these technologies will allow us to process data such as browsing behaviour or unique IDs on this site. Not consenting or withdrawing consent, may adversely affect certain features and functions.
Functional
Always active
The technical storage or access is strictly necessary for the legitimate purpose of enabling the use of a specific service explicitly requested by the subscriber or user, or for the sole purpose of carrying out the transmission of a communication over an electronic communications network.
Preferences
The technical storage or access is necessary for the legitimate purpose of storing preferences that are not requested by the subscriber or user.
Statistics
The technical storage or access that is used exclusively for statistical purposes.The technical storage or access that is used exclusively for anonymous statistical purposes. Without a subpoena, voluntary compliance on the part of your Internet Service Provider, or additional records from a third party, information stored or retrieved for this purpose alone cannot usually be used to identify you.
Marketing
The technical storage or access is required to create user profiles to send advertising, or to track the user on a website or across several websites for similar marketing purposes.
