Decision Making and AI for business change

Decision making

Process for:

• Identifying a difference between the present and the desired status
• Making choices in order to manage this difference (eliminate or reduce)

For some problems the process is formalized:

• Models
• General resolution methods
• Specific solving methods

Which problems

• Sufficiently complex and/or relevant
• The economic and analysis time effort must be justified

The decision making normally starts from the definition of the problem to the choice of an alternative.

Decision making: Where?

If you have problems which are complex or large sized or problems difficult to define, typically here you have AI.

Because it’s able to define the problem starting from the data. You have to formalize the data structure, it’s a databased approach. The counter effect problem is that a deep analysis is required before taking the decisions and the definition which are the decisions to take.

New problems (expertise not sufficient).

Automation and routine decisions.

Strong interaction between decision maker (management) and problem expert.

When you have decision making?

You can have strategic decisions, typically long term decision with big uncertainty. Then even tactical made of implementation of strategic decisions, affect the usage of resources, may be affected by uncertainty on the data, tactical lasts months or weeks (industry dependent).

Then you have the operational decisions, short term decisions and here uncertainty is normally limited, the time period is usually days or hours.

Then the real time, very short time for the decision (second or less), here the decision might be real-time for the application or due to the technology, blockchain required to take an optimized decision in short time to avoid performances issues. Another example is the e-commerce and the need of short-term decisions.

Qualitative is from experience and intuition of managers and practitioners, are problems relatively simple and small, the final results strongly related to experience and skills of the decision makers.

While quantitative require a set of specific knowledge and methods to be mastered (math, solution methods, etc.) It can deal with large problems, and need of specifically trained staff.

• Strategic: Stochastic programming, bilevel programming, exact approaches, metaheuristic approaches
• Tactical: Stochastic programming, metaheuristic approaches
• Operational: Combinatorial optimization (usually heuristics and metaheuristics), Artificial Intelligence
• Real-time: Artificial Intelligence, simple heuristics

It’s important to define of course the application domain, for example finance, forecasting, PM, productions, design, transportation, distribution, social networks, marketing etc.

Other names could be: Management Science, Operations Research, Operational Research, Artificial Intelligence, Business Intelligence, Analytics, Quantitative Methods, Decision aiding methods, Science of Better.

Models

We work by means of models. Arbitrary representations of a system object, organisation, processes etc. It transforms the information into decisions by means of (mathematical) models.

Don’t try think to solve the real issue, but to model and to give solution to the realistic system not to the real system.

We use models to understand and evaluate our environment and to follow its changes.

For management

• To have better knowledge of the system, its goals, constraints and parameters.
• Identify actors and blocks and their interactions, in order to predict system evolution.
• Characterize system problems.
• Manage (solve, hopefully) the identified problems.
• Test before implementing (scenario based simulations and optimization).

Usually is better to “play” with a representation that consider the real system because it’s quicker, less costly and less risks.

Example: Production planning example

A manufacturer produces a wide variety of bicycles. We are interested in the production plan of a single high-tech racing model whose production requires special materials and production equipment. At most one batch is produced per month, because of low demand and important economies of scale in the manufacturing costs. Because of the need to install special equipment and tools at the beginning of a batch, there is a high set-up cost, and thus it has been decided that it makes no sense to produce more frequently.

The batch manufacturing cost is approximated by the fixed charge cost. The set-up cost represents the equipment and tool installation and preparation costs, and then the constant marginal cost corresponds to the constant time required to produce each bicycle. For the racing model, the set-up cost is € 5000, and the marginal cost is € 100. Hence, it costs € 5100 to produce a batch of 1 bicycle, and € 6000 for a batch of 10 bicycles.

The capacity restrictions are ignored in planning this single-product variant because the work center and workers are shared by the many bicycle variants, and because capacity can be increased by hiring temporary workers, if necessary. The company faces irregular or seasonal demand, sales being higher in spring and summer. Table below gives the sales forecasts in number of bicycles per month for the racing model.

Moreover, there will be around 200 at the beginning of the next year. To hold one bicycle in inventory during one month costs on average € 5, including the cost of capital and storage costs. Note that there is enough space available in the warehouse to store the bicycles. The manufacturer wants to plan the production and inventory levels of this particular racing bicycle, in order to satisfy demand and minimize the corresponding manufacturing and inventory costs. He wants to plan production for next year up to the end of the peak demand period, that is, up to the end of August.

First typical solution

Because of the economies of scale, production costs are minimized by producing very large batches. Produce the whole demand.

But this is not the right solution, it pays a big inventory costs. In order to reduce it, you produce month by month.

Second typical solution

We try to optimize inventory and production costs at the same time. We produce every two months.

A not so intuitive solution.

Optimal solutions

The inventory cost must be less than the setup cost. Thus the storage is efficient up to 1000 units (why?). 1000 * 5 is 5000.

Another example

3 CPU: M1, M2, M3.

3 Job: Each job requires all the CPUs and have the same sequence M1, M2, M3.

When a job stops on one CPU is immediately dispatched on the other one. Parallel architecture of some specific hardware (Pipeline of multithread processors).

Flow Shop: operations of every job must be processed on the machines in the same order.

No-wait: when an operation on a machine stops, it must start on the other machine immediately.

Common constraint when there is no waiting areas for the jobs.

Suppose:

3 Jobs: J1, J2, J3.

3 production stages, 1 CPU per stage: M1, M2, M3.

Schedule of the jobs: M1, M2, M3.

Processing times

• J1: 2 2 6
• J2: 2 6 2
• J3: 6 2 2

The optimal schedule that you can find is this one. So, the makespan is 14.

A new problem

We buy a new machine for the second stage.

We use the second machine only.

The processing times of the new machine are half of the previous one.

Processing times

• J1: 2 1 6
• J2: 2 3 2
• J3: 6 1 2

Compute the new optimal solution. Hint: how many solutions can we have?

Optimal solutions

We have 6 permutations of the jobs:

• J1, J2, J3: Cmax 15
• J1, J3, J2: Cmax 15
• J2, J1, J3: Cmax 15
• J2, J3, J1: Cmax 17
• J3, J1, J2: Cmax 17
• J3, J2, J1: Cmax 19

Is there something weird?

Other variants of the FS paradox

Consider the same instance.

Cmax = 14.

Reduce the processing times of job J2 by 25% Cmax = 15.5.

Reduce the processing time of job J2 in machine M2 from 6 to 3 Cmax = 15.

But even doing so, it’s not working.

Graham’s multiprocessing anomaly concerns parallel machine systems where precedence-constrained processes are assigned according to the list-scheduling policy. By adding new machines to the system, one might actually worsen the makespan.

Belady’s anomaly

In common computer memory management, information is loaded in specific sized chunks. Each chunk is referred to as a page. The central processor can only load a limited number of pages at a time. It requires a frame for each page it can load. A page fault occurs when a page is not found, and might need to be loaded from disk into memory. When a page fault occurs and all frames are in use, one must be cleared to make room for the new page. A simple algorithm is FIFO. Whichever page has been in the frames the longest is the one that is cleared. The paradox: it is possible to have more page faults when increasing the number of page frames while using FIFO page replacement algorithm.

Models solutions

Predictive models

Predictive models: predict or estimate future parameters and/or events. Examples are forecasting of the rainfalls. Possible methods include regressions analysis (linear, logistic, etc.), time series forecasting, machine learning (neural networks, random forests) and ensemble methods (bagging, boosting).

Descriptive models

Descriptive models: describe and understand ongoing phenomena or processes (chatgpt). Examples are the identification of how users navigate a website, support radiological analysis by identifying patterns in medical images. Possible methods include descriptive statistics, data visualization, clustering, dimensionality reduction and sentiment analysis.

Prescriptive models

Prescriptive models: generate recommendations on possible actions or strategies. Examples include proposing patient-specific chemotherapy regimens, scheduling and packing/loading in the industry. Possible methods include optimization techniques, decision analysis, heuristics/metaheuristics methods and simulation-optimization (monte carlo simulations integrated with optimization).

The AI Value Chain Iceberg

Visible Surface: 10% User interfaces, chatbots, and generative AI applications represent only the visible tip of a vast systemic structure.

Hidden Depths: 90% Infrastructure, energy markets, capital allocation, and geopolitical forces constitute the submerged foundation, where systemic risks accumulate.

The $800 Billion Discrepancy Capital Mobilization at Scale

Big Tech has invested over $800 billion in AI infrastructure over five years, exceeding the GDP of midsized nations.

The 'Land Grab' Gap

Current revenue returns do not justify expenditure on traditional profit-and-loss metrics. This represents fundamental economic restructuring, not a product cycle.

Energy Deficit: Physical Cost of Virtual Intelligence

Projected Consumption by 2026: Data center energy usage will triple, reaching levels equivalent to a major G7 nation's total consumption.

Cooling Requirements: 40% of energy is consumed solely for cooling infrastructure. The cloud operates as heavy industrial machinery running 24/7.

Infrastructure Imperative: Physical demands require strategic policy coordination and energy innovation at national scale.

Privatized Gains, Socialized Costs

The Grid Externality Mechanism

Residential energy consumption remains constant, yet utility bills rise sharply as infrastructure upgrade costs are distributed to residents through regulatory loopholes.

Case Study: Northern Virginia & Ohio. Residents unintentionally subsidize Big Tech's operational expansion through grid regulation structures.

Structural Pattern: Corporate infrastructure demands create public cost burdens without corresponding public benefit distribution.

Resource Competition: Agriculture vs. Algorithms

Water Scarcity Dynamics: Data centers in drought zones compete directly with agricultural and civil water use for cooling requirements. One facility consumes 1,000,000 gallons of water per day.

Governance Question: Who creates the priority list? Farms or Server Farms?

Desperate Measures for Power

Nuclear Reactivation

Microsoft & Constellation Energy restarting Three Mile Island. Amazon utilizing Susquehanna plant for data center power.

Underwater Centers

China's strategy off Shanghai coast. Ocean cooling saves 90% energy but faces corrosion and maintenance risks.

Orbital 'Moonshots'

Google's 'Sun Catcher' concept. Space-based data centers for vacuum cooling and solar power. Technically distant but signals terrestrial scarcity (and what about Musk?).

The Circular Economy of AI Value

Round-Tripping Mechanism

• NVIDIA invests capital in cloud providers
• Cloud providers rent compute to Big Tech
• Big Tech purchases chips exclusively from NVIDIA
• Stock prices rise on circular revenue flows

Question: Does this structure represent genuine value creation or financial engineering?

Systemic Risk: The Central Node

Single Point of Failure

280% growth rate for NVIDIA's valuation increases over two years and $3T of market capitalization exceeding most national economies. “if one tile falls, the tower collapses”, a structural mimicry of historical financial bubbles.

The Erosion of State Sovereignty

The Structural Shift

Corporations transitioning from lobbying government to staffing government institutions.

• Tech CEOs holding government efficiency roles
• Private entities defining National Defense standards (SpaceX, Palantir)
• Infrastructure ownership creating regulatory capture

Can a state regulate entities that own its critical intelligence infrastructure?

Geopolitics: The Asymmetric War

Strategic Degradation Result: A race where speed is prioritized over safety protocols. Strategy focuses not on advancement but on degrading opponent's reality perception.

Poisoning the Well: Data contamination through false positives flooding training datasets.

Algorithm Confusion: Exploiting logic gaps in decision-making systems.

The Four Pillars of Governance Necessity

Environmental: Manage externalities (water/energy). Protect public resources from private extraction.

Economic: Prevent circular valuation bubbles and monopolistic concentration. Ensure fair competition.

Societal: Protect democratic agency. Prevent behavioral manipulation at population scale.

Geopolitical: Establish national security standards. Prevent critical infrastructure failure.

Why This Course Matters

To effectively govern today's complex systems, individuals must deeply understand both the opportunities and the inherent challenges of AI. What is readily apparent is often just a small fraction of a much larger, intricate system.

We will analyze AI from several crucial perspectives:

• AI as Part of Decision Science Understanding AI within a broader technical framework focused on informed decision making.
• AI as an Integrable Technology Exploring how AI fits into and enhances existing systems and infrastructures.
• AI as a System to Govern Addressing the critical need for comprehensive frameworks to manage and regulate AI's impact.

The LETO Methodology

A comprehensive framework for governance, integrating critical dimensions into a cohesive approach.

Legal: Establishing clear regulatory frameworks, compliance, data privacy laws, and accountability mechanisms for AI systems.

Ethics: Ensuring fairness, transparency, human oversight, and addressing societal impacts including environmental and social considerations.

Technology: Focusing on the secure development, reliable implementation, and technical safeguards within AI design and deployment.

Operational: Defining best practices for deployment, monitoring, maintenance, and responsible use of AI in real-world scenarios.

Linear Programming Models

Here, we are in the point of model construction.

Why we need models?

Define the (computer) complexity of a problem ⇒ Solve to optimality with specific solvers ⇒ Define the properties of a problem ⇒

To identify the decisions we know that the decisions are represented by decision variables.

Clarify the objective of the problem that you have to express the objective through one (or more) objective function. Then to identify the constraints of the problem and express them through combinations of the variables.

Linear Programming is the part where the objective function and the constraints are linear combinations of the variables.

LP – Linear Problem is the when all variables are continuous while MIP stands for mixed integer problem (when some of the variables are integral, even one). But the most one are made of linear combination.

Looking at the example of yesterday

Problem.

Decision makers:

• Chief Production Officer

Planning of the optimal production ⇒ What-if analysis based on updated forecasts of the demand ⇒

Information.

Production periods: 8 months – index j ⇒ Demand: Highly fluctuating ⇒ No detailed information on demand variability ⇒ Hypothesis: we assume the demand as known ⇒

Which are the decisions?

Production: Setup: Warehouse: The warehouse level is actually a decision? No: it is set as a consequence of production, demand, and warehouse level at the previous period.

= + – −1 j

Constraints

Problem:

• Demand must be satisfied
• Warehouse must be empty at the end of the period (end of August)
• Set up cost must be considered only if there is production

Solution Homework ⇒

I can produce at maximum 7000.

Exercise 1: A production planning problem

Svivon Company produces three types (Alfa, Beta, Gamma) of electric batteries. For the production of two types (Beta and Gamma) uses copper. To satisfy the production for the next month, it is possible to buy copper at the price of 5 euro/kg. The supplier can supply no more than 4000 kg of copper. The following table shows: the quantity of copper needed for each type of battery, the production costs (for each battery) and the selling prices (for each battery).

The number of Alfa batteries must be at least twice the number of Beta batteries. The number of Alfa batteries must not be higher than the number of Gamma batteries. Formulate the LP model for the opti