Be it in factories, asset fleet management, intra-logistics or supply chain management – material flow has to be effectively controlled, optimized and maintained. Whereas the human brain can solve these challenges with practice and unparalleled creativity in smaller processes, it meets boundaries in on-scale processes: Leveraging IoT sensors, cloud infrastructure and AI algorithms enable efficient data-driven automated decision making for logistics at scale, resulting in reduced storage cost, more effective asset utilization and more effective just-in-time delivery. In fast-paced and large logistic processes every percent of effectiveness leveraged yields high business value.
Saving resources with data-driven decisions
Our customers have world-wide supply chains for their production plants. To ensure materials are in stock when needed in order to avoid costly production stops, supply chain logistics are planned ahead of time.
Most of the necessary freight is carried by cheap but slow sea travel. However, many events can change the situation while a plan is executed. Parts may be damaged or missing, strikes may stop parts of the supply chain, weather events may delay scheduled arrivals. There are countless possibilities for the original plan not to work out as expected.
To ensure uninterrupted production, our customers’ supply chain planners decide on sending fast but expensive air freight based on available stock and supply chain data on a daily basis. Historically, these decisions have been made based on experience and intuition of the planners. AI supports the planners’ intuition by determining the need for air freight algorithmically.
Deciding whether to send air freight or not is a challenge recurring very frequently. By increasing the quality of this decision just by a fraction, our data-driven algorithmic solution is able to reduce air freight cost by millions per year, let alone reduction of CO2 emissions.
Every decision matters
Should we send freight by air or by sea? When to send it? On which asset should cargo be loaded? Where should this truck go to ensure maximum future utility? When should it go? Does this asset need maintenance now or later? When to schedule maintenance to maximize utility?
In logistics and supply chains many small decisions are taken at a high frequency. Making these decisions as effective as possible is of crucial importance. Reducing the cost of a single decision by one percent may not sound spectacular. If you make this decision a million times a year, however… you get the point. Every percent matters!
The fundamental components of good decision making in logistic processes
Making good decisions in a large-scale process has three fundamental components: Being informed about the present state of the process, being able to estimate future dynamics for making the best informed decisions possible, and being able to take and combine decisions for achieving the best outcome in the logistic process.
Being informed about the process: Machines are able to see and communicate at higher resolution and frequency than humans. As long as humans are involved in decision making and controlling, it is necessary to provide human-interpretable abstractions of what’s going on. Well-chosen visualizations may help making business processes more transparent and interpretable for decision makers.
Consequences and automated decision making: Leveraging simulation (a digital twin) enables to automatically evaluate many different “what-if” scenarios for choosing the best actions in a particular situation. Predictive modeling (ETA, demand distribution, predictive maintenance) allows to incorporate real observations (data) from the business process into the simulation, increasing its accuracy and consequently the quality of decisions.
Combining decisions for best outcome: In a logistics process, each decision impacts many other decisions that can be made in parallel or in the future. For example, deciding on a truck driving to Hamburg will impact the choices for routing this truck on its next trip – Rome is far away, however Berlin may be calling. Also, other trucks may not have to serve Hamburg because of this decision and are free to be distributed to serve different needs. This interleaving of decisions’ impact makes it hard for humans to make the best possible decision, even given they have all relevant information – the process itself is too complex to be analyzed for decision making in feasible time by a human. Here, data driven algorithmic decision making can be leveraged to increase the quality of decisions.
Why isn’t everybody using it? Are there reasons not to do it?
Infrastructure and non-trivial application landscapes: Putting AI models to production requires a substantial amount of software engineering and enterprise architecture activities. Without those, its hard to quantify the potential RoI of any AI application. Without those quantities however decision makers may be reluctant to invest.
Real-time data and data quality: Real-time decisions require real-time information about the process to be controlled. The data collection and harmonization infrastructure may not be in place yet, creating a non-trivial hurdle to bypass for any efforts towards AI-based logistics. For example, fleet containers may need to be equipped with localization sensors, or backend infrastructure needs to be built to gather and process real-time GPS data from trucks in a logistics fleet.
Lack of knowledge and skills to map business processes and AI application possibilities: A deep understanding of both business processes and technological possibilities and limitations is required to identify core ingredients of AI-driven supply chain automation and optimization. Modeling the process, acquiring a useful simulation and connecting to the optimization tools is a highly non-trivial task.
PoC effort and RoI timescales are over-estimated: While identification of relevant processes and algorithms is a complex undertaking, the effort needed to implement a PoC and testing the viability of an idea is remarkably low, assuming infrastructure is already in place. Many initial cases can be tackled with pretty simple algorithms, immediately providing quantifiable information about RoI.
Collaboration of business units is a necessity: Business and IT departments have to collaborate closely to identify relevant applications and actions to be taken for realizing their potential, such as adaptation of processes and innovation to IT infrastructure and digital representation of running business processes.
Innovation vs. daily business: AI adoption is still a young trend. Innovation activities compete with daily business needs and maintenance for budget and resources, and often there is limited budget for innovation activities.
R&D and risk: Research and development is characterized by high risk activities. Failure has to be an option. In many cases the business potential of an approach cannot be quantified before performing a first analysis of available data. Many decision makers avoid taking these risks. As there are few first-movers for AI in logistics, aspirants may choose to wait for others to succeed before starting their own activities.
Testing & safety: While quality assurance is technically possible for AI systems, it is a subtle and complex undertaking. Many AI components fail silently, i.e. error only become visible over time or in a statistical fashion. Deep knowledge of AI dynamics in production and available algorithmic solutions is necessary to assure quality of AI components.
Trust & control: Humans want to stay in control and tend to consider their cognitive abilities superior to any machine – “that’s impossible” is a common reaction when showing effective AI decision making to experienced practitioners.
Most of the AI/ML activities we observe in business context are R&D undertakings, with a focus on experimentation and iteration. Some use cases are realized as PoCs to evaluate their potential in real business settings.
In cases we have worked on we see promising results. Predictive capabilities (for example predicting energy consumption, estimated arrival times or asset idle times) are typically increased in two-digit percent ranges regarding prediction error compared to currently employed heuristics or baselines. The same holds true for automated decision making and process control, where we observed up to 7-digit yearly saving potential based on evaluation on real data.
Microsoft AI Award with RAIL: We won the Microsoft AI Award 2019 with our proof-of-concept for supply chain optimization with automated planning. Our system optimizes supply chain control more effectively than heuristic approaches and classical operations research algorithms by coordinating the whole asset fleet on a global scale in real-time and robust to unexpected events.
Network Load Balancing: Currently, rail network load balancing is based on intuition and experience of human operators. We support them with data driven analysis, and increase effectiveness of load balancing in a railway cargo transportation network by identification and prediction of network capacity overflow.
Supply Chain Optimization: Deciding whether to send goods via ship or airplane is at the core of this project. We automate short-term planning and decision making for supply-chain control on a global scale, saving millions per year by real-time planning.