Augmented Intelligence for transport planning: Human in the Loop Modelling

An optimization model learns to see snow instead of a wolf. This is why even the best algorithms still need a human in the loop.

#1about 3 minutes

Applying augmented intelligence to logistics planning

Decision support systems can solve complex logistics problems by combining statistical estimation and operational research.

#2about 6 minutes

Understanding different mathematical optimization techniques

Techniques like linear, mixed-integer, and convex programming provide a powerful declarative way to model and solve business problems.

#3about 4 minutes

Why human planners often distrust optimization models

Planners may reject optimized solutions due to unstated constraints, responsibility concerns, and a mismatch between global optimization and local decision-making.

#4about 4 minutes

How models learn incorrect correlations from data

Examples like classifying wolves by snow in the background show how models can achieve high accuracy for the wrong reasons, justifying user skepticism.

#5about 3 minutes

Declarative models versus imperative human thinking

Optimization models require full commitment to a declarative plan, whereas humans prefer an imperative, step-by-step approach with partial execution and replanning.

#6about 2 minutes

A human-in-the-loop framework for building trust

A proposed workflow involves generating suggestions, allowing planners to review and provide final decisions, and giving them tools to examine and modify results.

#7about 7 minutes

Using probabilistic programming for incomplete data

Probabilistic programming helps estimate missing package dimensions by modeling the data generation process, allowing for flexible handling of missing values and incorporating expert knowledge.

#8about 5 minutes

Techniques for model interpretability and transparency

Building user trust involves making models more transparent by extracting human-readable rules and clearly visualizing the uncertainty in their predictions.

#9about 8 minutes

Interactive scenario planning and constraint editing

Allowing users to collaboratively build models through scenario analysis and then interactively edit constraints in real-time bridges the gap between optimization and practical operations.

#10about 3 minutes

Conclusion: Achieving symbiosis between humans and AI

The key to successful adoption is creating a symbiotic relationship where expert users can guide and interact with optimization models through thoughtful UI and solution presentation.