Bayesian Probabilistic CPM

• Bayesian Probabilistic CPM is a dynamic scheduling method that models activity durations as random variables updated with real-time data.
• It employs Monte Carlo simulation to generate finish-date CDFs and calculate risk percentiles (e.g., P50, P80) along with activity criticality indices.
• Integrating data from drone photogrammetry, LiDAR, and computer vision, it improves responsiveness and risk control in construction projects.

Bayesian Probabilistic CPM is a probabilistic generalization of the Critical Path Method (CPM), employing Bayesian inference and Monte Carlo simulation to dynamically update activity durations, project completion forecasts, and criticality indices based on real-time progress data. It is designed to overcome the inherent inflexibility and deterministic limitations of classical CPM scheduling, especially under uncertainty and in contexts with frequent field updates. The framework integrates real-time observational data streams (e.g., drone photogrammetry, LiDAR, computer vision-derived percent-completes) with prior distributions for activity durations, enabling transparent, responsive, and risk-aware project control. The Bayesian CPM system operationalizes schedule risk metrics—including finish-date uncertainty bands (P50, P80), buffer consumption, and critical activity probabilities—making it suitable for advanced construction project management and digital-twin environments (Khoshkonesh et al., 4 Nov 2025).

1. Bayesian Modeling of CPM Activity Durations

Each schedule activity $i$ ($i = 1,\dots,N$) is assigned a total duration $D_i$ treated as a random variable with prior $\pi(D_i)$ reflecting planner beliefs or historical skew. Two principal prior forms are used:

• Triangular prior: If planners provide optimistic, most-likely, and pessimistic bounds $(a_i, c_i, b_i)$, then $\pi(D_i) = \mathrm{Triangular}(a_i, c_i, b_i)$.
• Log-normal prior: For right-skewed historical durations, $D_i \sim \mathrm{LogNormal}(\mu_i, \sigma_i^2)$.

Upon each weekly update $k$, site-based percent-complete measurements $m_{i,k}$ are observed through scan-to-BIM work breakdown reconciliations and computer vision status signals. The likelihood model is:

$$m_{i,k} \sim \mathrm{Normal}\left(\frac{t_{i,k}}{D_i}, \sigma_\epsilon^2\right)$$

where $t_{i,k}$ is cumulative elapsed time for activity $i$ at update $k$. The joint posterior after $K$ updates integrates sequential observations:

$$p(D_i \mid m_{i,1:K}) \propto \pi(D_i) \prod_{k=1}^K L(m_{i,k} \mid D_i)$$

with the normalization performed over the support of $D_i$.

2. Probabilistic Schedule Aggregation via Monte Carlo Simulation

Once posteriors $p(D_i | m_{i,1:K})$ are obtained for all activities, probabilistic schedule metrics are computed using Monte Carlo methods. For each of $S$ simulations:

1. Sample $D_i^{(s)} \sim p(D_i)$ for all activities.
2. Compute $$T^{(s)} = \operatorname{LongestPathLength}(G, \{D_i^{(s)}\})$$, where $G$ is the CPM precedence graph.

The ensemble $\{T^{(s)}\}_{s=1}^S$ forms an empirical CDF:

$$\widehat{F}_T(t) = \frac{1}{S} \sum_{s=1}^{S} \mathbf{1}\{T^{(s)} \leq t\}$$

From which risk percentiles (e.g., P50 median, P80 conservative forecast) and activity-level criticality indices are computed:

$$\text{Crit}_i = \frac{1}{S} \sum_{s=1}^{S} \mathbf{1}\{i \in \text{any longest path in sample }s\}$$

3. Data Assimilation and Real-Time Bayesian Updating

The framework systematically fuses heterogeneous field data:

• Weekly drone photogrammetry and monthly LiDAR—producing scan-to-BIM percent complete, $m_{i,k}^{(\text{scan})}$.
• Site imagery (CV-derived)—yielding status signals, $m_{i,k}^{(\text{CV})}$.

Updates occur on a rolling-window basis:

• Combine $m_{i,k}^{(\text{scan})}$ and $m_{i,k}^{(\text{CV})}$ for unified likelihood input.
• Refresh the posterior $p(D_i| \text{data})$ after each round.
• BayesUpdate is executed using the latest percent-complete and elapsed time for each $i$.

Each update recalibrates the probabilistic schedule dashboard: finish-date CDF, buffer metrics, activity criticalities.

4. Algorithmic Workflow

The complete scheduling control logic is as follows:

// Initialization for each activity i: set prior π(D_i) (triangular or log-normal) load precedence graph G for each weekly update k = 1..K: ingest scan-to-BIM percent completes {m_{i,k}^{scan}} ingest CV status signals {m_{i,k}^{CV}} for each i: combine m_{i,k}^{scan}, m_{i,k}^{CV} into m_{i,k} update posterior p(D_i) ← BayesUpdate(π(D_i), m_{i,1:k}, t_{i,1:k}) // Monte Carlo sampling for s = 1..S: for each i: sample D_i^{(s)} ∼ p(D_i) T^{(s)} ← LongestPathLength(G, {D_i^{(s)}}) record activities critical in sample s compute finish-date CDF, extract P50, P80, etc. compute criticality_index_i = (times i is critical)/S compute buffer consumption for critical path activities output dashboard metrics end for

5. Empirical Performance and Numerical Results

In the Texas mid-rise case study (Khoshkonesh et al., 4 Nov 2025):

• Bayesian priors on activity durations updated weekly; after 13 weeks, the P50 finish-date forecast converged to 128 days, matching the actual completion (zero absolute error at close-out).
• Conservative P80 forecast plateaued at ≈130 days from week 9 onward.
• Buffer consumption tracked: feeding-buffer usage was 8/15 days; project-buffer usage 6/20 days.
• Highest criticality activities were envelope components (curtainwall/windows at 46%) and superstructure (post-tensioned slabs at 41%).

This demonstrates marked improvements over deterministic CPM—providing 43% labor reduction in cost estimating, 6% overtime reduction, and real-time detection of high-risk activities for targeted management.

6. Conceptual and Operational Significance

Bayesian Probabilistic CPM embodies data-driven scheduling wherein uncertainty is modeled explicitly and updated as field evidence accumulates. By integrating Bayesian inference, real-time measurement, and MCS, the method enables practitioners to:

• Track schedule risk and finish-date uncertainty in the presence of evolving field data and dynamic conditions.
• Quantify buffer consumption and activity criticality for granular risk control.
• Localize schedule interventions precisely where risk is concentrated.
• Ensure that schedule forecasts (e.g., P50, P80) are inherently responsive to stochastic field events and measurement noise rather than static baseline assumptions.

The approach is extensible to any project control environment where activity durations and dependencies can be captured, and progress data streams assimilated, supporting robust schedule management under uncertainty and across highly variable construction contexts.