Chapter 3: Real World AI and Uncertainty (Part 1)

Probabilistic Reasoning in a Noisy World

While game boards offer perfect information and deterministic rules, the real world is messy, noisy, and unpredictable. Sensors fail, information is missing, and outcomes are never guaranteed. Therefore, modern real-world AI abandoned rigid, rule-based logic in favor of Probability Theory.

Living in the arid regions of Marwar has always required reasoning under uncertainty. Agriculture and pastoral life depend entirely on Meh (monsoon rain), which rarely arrives on a fixed schedule. Farmers and shepherds do not paralyze themselves waiting for absolute certainty; they express expectations through odds and probabilities. We teach learners how to convert comfortably between odds (such as 3:1 in favor of rain) and percentages (75% probability), establishing the mathematical foundation for handling imperfect knowledge.

Updating Beliefs with Bayes Rule

How should a rational thinker or an algorithm alter its beliefs when new evidence comes to light? The answer lies in Bayes Rule, a cornerstone of modern AI. The rule provides a mathematically rigorous way to combine prior beliefs with new observations to calculate an updated posterior probability:

Posterior Odds = Likelihood Ratio × Prior Odds

We demonstrate this principle through medical screening, explaining the Base-Rate Fallacy: the human tendency to focus solely on a positive test result while ignoring the overall rarity of the disease in the broader population.

To ground this locally, consider a Marwari farmer predicting rain ahead of the Teej or Gangaur festival. Based on the historical Choumaso calendar, the farmer starts with baseline Prior Odds of rainfall. Suddenly, they observe new evidence: a sudden change in wind direction accompanied by the dark clouds of a Sunto storm. The farmer evaluates the Likelihood Ratio (how much more likely this storm pattern is to occur when rain is imminent compared to when it is a dry dust storm). Multiplying the prior odds by this ratio yields sharp Posterior Odds. This belief update governs whether the village proceeds with Halsotiyo (the first plowing ritual) or holds back seed grain in the Khalo.

Naive Bayes and Text Classification

We apply probabilistic reasoning to build practical tools like email spam filters using the Naive Bayes Classifier. When analyzing a message, the algorithm evaluates the words present and calculates the combined odds that the message is junk versus legitimate correspondence.

The system is called "naive" because it makes a simplifying assumption: it treats every word as conditionally independent of the others, ignoring grammar and word order. Despite this simplification, it operates with remarkable accuracy and computational speed. This practical rule of thumb is similar to how a young boy taking his cattle out to graze quickly classifies whether a patch of land is fertile pasture or barren ground. Instead of conducting a complex soil analysis, he evaluates independent clues: the presence of specific thorny bushes, the types of dry grasses, or the surface moisture. By weighing these individual visual markers independently, he arrives at a fast and reliable classification of the terrain.