2.6 Decision Trees, Expected Value & Bottleneck Analysis (Theory of Constraints)

2.6 Decision Trees, Expected Value & Bottleneck Analysis (Theory of Constraints)

Industrial and systems engineers must routinely make high-stakes capital, capacity, and operational choices in the presence of uncertainty, market volatility, and operational constraints. This section synthesizes quantitative decision theory—spanning criteria under uncertainty, Bayesian decision trees, and value-of-information calculations—with Eliyahu M. Goldratt's operational Theory of Constraints (TOC) and bottleneck management.

---

Decision Theory Taxonomies & Payoff Matrix Formulation

A formal decision problem consists of three mathematical structures:

1. Decision Alternatives ($d_i \in D$): A mutually exclusive, collectively exhaustive set of $m$ candidate actions controlled directly by the decision maker (e.g., build large plant, build small plant, do nothing).
2. States of Nature ($s_j \in S$): A mutually exclusive, collectively exhaustive set of $n$ uncontrollable future environmental conditions or market outcomes (e.g., high demand, moderate demand, low demand).
3. Payoff Outcomes ($v_{ij} = V(d_i, s_j)$): The quantifiable consequence (profit, contribution margin, net present value, or cost) realized if alternative $d_i$ is chosen and state of nature $s_j$ subsequently occurs.

The Payoff Matrix

Three Decision Environments

• Decision Making Under Certainty: The future state of nature is known with 100% deterministic certainty ($P(s_k) = 1.0$). Solved by straightforward deterministic optimization.
• Decision Making Under Risk: The decision maker cannot predict the exact state of nature, but possesses reliable objective historical data or subjective engineering estimates to assign a probability distribution $P(s_j)$ across all states, where $\sum_{j=1}^n P(s_j) = 1.0$.
• Decision Making Under Uncertainty: The states of nature are known, but no probability distribution can be assigned to their occurrence.

---

Decision Making Under Uncertainty (Non-Probabilistic Criteria)

When state probabilities are completely unknown, engineers apply four classical criteria depending on managerial risk tolerance:

1. Maximax Criterion (Optimistic / Extreme Risk-Seeking)

Select the decision alternative that maximizes the maximum possible payoff:

$d^* = \arg\max_i \left( \max_j v_{ij} \right)$

2. Maximin Criterion (Wald's Conservative / Risk-Averse)

For each alternative, identify the worst-case (minimum) payoff, then select the alternative that maximizes this minimum guarantee:

$d^* = \arg\max_i \left( \min_j v_{ij} \right)$

3. Hurwicz Criterion (Criterion of Realism / Compromise)

Balances optimism and pessimism using an index of optimism $\alpha \in [0, 1]$. For each alternative, calculate the weighted payoff metric $H_i$:

$H_i = \alpha \cdot \left( \max_j v_{ij} \right) + (1 - \alpha) \cdot \left( \min_j v_{ij} \right)$ $d^* = \arg\max_i H_i$

When $\alpha = 1$, Hurwicz equals Maximax; when $\alpha = 0$, Hurwicz equals Maximin.

4. Minimax Regret Criterion (Savage Opportunity Loss)

Focuses on avoiding post-decision regret. The regret (opportunity loss) $R_{ij}$ is the difference between the best possible payoff for state $s_j$ and the actual payoff achieved by choosing alternative $d_i$:

$R_{ij} = \left( \max_k v_{kj} \right) - v_{ij} \quad (\text{for profit payoffs})$

Construct the Regret Matrix $[R_{ij}]$, determine the maximum regret for each alternative $\max_j R_{ij}$, and select the alternative that minimizes this maximum exposure:

$d^* = \arg\min_i \left( \max_j R_{ij} \right)$

Comparative Numerical Demonstration

Consider an industrial manufacturer deciding between three capacity expansion strategies under three market demand scenarios (payoffs in millions of dollars):

• Maximax Selection: $d_1$ (Large Facility) with maximum payoff $100M.
• Maximin Selection: $d_3$ (Small Facility) with guaranteed minimum payoff $20M.
• Hurwicz Selection ($\alpha = 0.6$): $d_1$ with score $0.6(100) + 0.4(-20) = \text{\textdollar}52\text{M}$.

Opportunity Loss (Regret) Matrix Construction:

• For $s_1$: Best is $20M. $R_{11} = 20 - (-20) = 40$; $R_{21} = 20 - 10 = 10$; $R_{31} = 20 - 20 = 0$.
• For $s_2$: Best is $40M. $R_{12} = 40 - 40 = 0$; $R_{22} = 40 - 30 = 10$; $R_{32} = 40 - 25 = 15$.
• For $s_3$: Best is $100M. $R_{13} = 100 - 100 = 0$; $R_{23} = 100 - 60 = 40$; $R_{33} = 100 - 30 = 70$.

Minimax Regret selects $d_1$ or $d_2$ with maximum regret $40M.

---

Decision Making Under Risk: EMV, EVPI & EVSI

When prior probabilities $P(s_j)$ are available, probabilistic expected value models guide rational choices.

1. Expected Monetary Value (EMV)

The expected return of alternative $d_i$ is the probability-weighted sum of its payoffs across all states:

$EMV(d_i) = \sum_{j=1}^n P(s_j) \cdot v_{ij}$

The rational decision maker selects the alternative with the highest expected value: $EMV^* = \max_i EMV(d_i)$

2. Expected Value of Perfect Information (EVPI)

Suppose an infallible market research agency could provide a 100% accurate forecast of the future state of nature before the decision is made. How much should an organization pay for this perfect intelligence?

• Expected Payoff Under Perfect Information (EPPI / EVWPI): If the state is known in advance, management will always select the best alternative for that specific state:

$EPPI = \sum_{j=1}^n P(s_j) \cdot \left( \max_i v_{ij} \right)$

• Expected Value of Perfect Information (EVPI): The difference between the expected return with perfect information and the maximum expected return without information:

$EVPI = EPPI - \max_i EMV(d_i)$

Fundamental Identity: EVPI is mathematically identical to the Expected Opportunity Loss (EOL) of the optimal EMV decision: $EVPI = EOL^* = \sum_{j=1}^n P(s_j) \cdot R_{i^* j}$

3. Expected Value of Sample Information (EVSI)

In reality, market research or prototype testing is imperfect. The Expected Value of Sample Information (EVSI) measures the economic value of sample testing:

$EVSI = EMV_{\text{with sample info}} - EMV_{\text{prior without info}}$

• Efficiency of Sample Information ($\eta$): $\eta = \frac{EVSI}{EVPI} \times 100\%$

---

Decision Tree Modeling & Bayesian Probability Revision

A Decision Tree is a chronological graphical structure that maps complex sequential decision alternatives and uncertain chance events.

Decision Tree Syntax

• Decision Node (Square $\square$): Represents an active point of managerial control. Branches emanating from a decision node represent mutually exclusive alternatives. Evaluated by selecting the maximum branch value (or minimum for costs).
• Chance Node (Circle $\bigcirc$): Represents an uncertain environmental event. Branches represent possible states of nature, each labeled with its conditional or marginal probability. Evaluated by calculating the expected value (EMV) across all outgoing branches.
• Terminal Node (Triangle $\triangle$ or End Point): Represents the final cumulative monetary outcome associated with that specific decision-state path.

Loading diagram...

The Backward Induction (Fold-Back) Algorithm

1. Draw the tree from left to right in chronological order of occurrence.
2. Compute the terminal monetary payoffs at each endpoint.
3. Roll back from right to left:
   • At each chance node, calculate the expected value: $EMV = \sum p_k V_k$.
   • At each decision node, select the branch with the highest EMV (pruning inferior branches).
4. Continue rolling back until the root decision node is reached.

Bayesian Probability Revision for Imperfect Information

When an imperfect test or inspection report $I_k$ is obtained, prior probabilities $P(s_j)$ must be updated to posterior probabilities $P(s_j \mid I_k)$ using Bayes' Theorem:

$P(s_j \mid I_k) = \frac{P(I_k \mid s_j) \cdot P(s_j)}{P(I_k)} = \frac{P(I_k \mid s_j) \cdot P(s_j)}{\sum_{m=1}^n P(I_k \mid s_m) \cdot P(s_m)}$

Where:

• $P(s_j)$: Prior probability of state $s_j$.
• $P(I_k \mid s_j)$: Reliability / likelihood of test indicator $I_k$ given true state $s_j$.
• $P(I_k)$: Total marginal probability of observing test result $I_k$.
• $P(s_j \mid I_k)$: Posterior probability that state $s_j$ will occur given observed indicator $I_k$.

---

Theory of Constraints (TOC) & Bottleneck Analytics

Developed by Dr. Eliyahu M. Goldratt in his seminal work The Goal, the Theory of Constraints (TOC) is a systems-management philosophy stating that every operational system is constrained by at least one bottleneck that dictates its maximum throughput.

Operational Classification of Manufacturing Resources

1. Bottleneck Resource: Any resource whose available capacity is equal to or less than the demand placed upon it. The bottleneck governs total plant throughput.
2. Non-Bottleneck Resource: Any resource whose capacity exceeds market demand. Running a non-bottleneck at 100% capacity does not increase throughput; it merely creates excess work-in-process (WIP) inventory.
3. Capacity-Constrained Resource (CCR): A resource that possesses adequate average capacity, but could become a temporary bottleneck if its schedule is unmanaged or if variability causes wave-loading.

Goldratt's Five Focusing Steps

Loading diagram...

1. Identify the Constraint: Find the physical or policy bottleneck. In a factory, it is easily identified by the largest accumulation of WIP inventory waiting in front of a workcenter and 100% equipment utilization.
2. Exploit the Constraint: Ensure 100% of the bottleneck's available capacity is dedicated to productive throughput:
   • Never allow the bottleneck to sit idle during lunch breaks, shift changes, or preventive maintenance.
   • Perform rigorous quality inspection before the bottleneck so that defective parts never consume scarce bottleneck capacity.
   • Offload non-critical tasks from the bottleneck to non-bottleneck machines.
3. Subordinate Everything Else to the Constraint: Align all upstream and downstream non-bottleneck resources to the exact processing rate of the bottleneck. Producing faster than the bottleneck generates harmful WIP, increases lead times, and clutters floor space.
4. Elevate the Constraint: If system throughput is still insufficient after exploiting and subordinating, invest capital to expand capacity (buy an additional machine, cross-train technicians, outsource operations).
5. Repeat (Prevent Inertia): Once the constraint is elevated, it will shift to another process. Return immediately to Step 1. Do not allow organizational inertia to perpetuate obsolete policies.

Drum-Buffer-Rope (DBR) Scheduling Methodology

DBR is the shop-floor execution mechanism derived from TOC:

• Drum (Pace): The bottleneck resource is the drum. Its operational schedule sets the beat/pace for the entire manufacturing facility.
• Buffer (Protection): A strategically placed buffer—measured in time rather than inventory piece counts—positioned immediately upstream of the bottleneck to protect it from upstream starving caused by minor breakdowns, setup delays, or operator absence.
• Rope (Synchronization): A formal pull communication link from the bottleneck to the gateway workcenter releasing raw materials into the plant. Raw material is released only when the bottleneck consumes an equivalent quantity, capping total factory WIP.

Throughput Accounting Metrics

TOC replaces traditional cost-accounting absorption methods with Throughput Accounting:

• Throughput ($T$): The rate at which the system generates money through sales: $T = \text{Sales Revenue} - \text{Totally Variable Costs (Raw Materials, Subcontracting)}$
• Investment / Inventory ($I$): All money tied up in the system (raw materials, WIP, finished goods, equipment, land).
• Operating Expense ($OE$): All money spent turning Inventory into Throughput (labor, rent, electricity, maintenance, depreciation): $\text{Net Profit} = T - OE$ $\text{Return on Investment (ROI)} = \frac{T - OE}{I}$

Little's Law in Bottleneck Analytics

Let $r_b$ denote the bottleneck processing capacity rate (units/hour), and $T_0$ denote the total raw process time through the line (hours). Under Little's Law:

$\text{WIP} = \text{Throughput} \times \text{Cycle Time} \implies W = \frac{L}{\lambda}$

• Critical WIP ($W_0$): The minimum WIP required to achieve maximum line throughput ($r_b$) with minimum theoretical cycle time ($T_0$):

$W_0 = r_b \cdot T_0$

If $\text{WIP} < W_0$, the bottleneck will starve and system throughput drops below $r_b$. If $\text{WIP} > W_0$, throughput remains capped at $r_b$ while cycle times and inventory holding costs increase linearly.