Buzzwords you would want to look up on what Wodinoneeye said are Bayesian Inference, Markov Chains, and MCTS (Monte Carlo Tree Search). There is a ton of work being done with the latter. Been some great presentations at AIIDE on them.

A good book you may want to at least look at :

Multi-Agent Systems (An Introduction to Distributed Artificial Intelligence)

by Jacques Ferber , 1999. Addison-Wesley

Has a good formalization of interactions between entities (agents)

Because you have to not just make your project work, but also present it in an academic manner.

The game/simulation is the problem space and a visualization system, the NPC AI part is the primary interest

You could implement a planner that builds schedules for daily tasks for a set of NPCs in a simplified symbolic environment.

The AI functions that decide fitness of decisions - like order to do them in or depletion priority (which thing of most importance I will run out of and need to get first). Time would be the cost there, different sources might have different costs. Having a time limit per daily 'run' as a ceiling limitation. Minimum requirements per day (like food) as a floor limitation. Some resources may have only one source and others more than one(with partial supply possible - requiring you to go to more than one. Some resources might (to complicate things) have a probability of their supply available - introducing uncertainty into the operation).

Goal is to minimize time and cost spent (on these routne tasks) leaving the remainder for profit (in that day)

Basic pathfinding on a map (to get distances between points of activities (obtaining/using resource) would be a 'tool'.

A Planner to decide the different sets of options (tasks in what order at which location for the days schedule)

Display of the results would be via some existing game library (someone mentioned Unreal4) to symbolicly represent objects and positioning.

Learning might come in where the results are statistically recorded day after day (need a result metric system for that) and compared to see which rule tunings of the AI decisions were the most effective

Swarm Logic (a method of monte carlo) possibly to adjust/supply the tuning candidates

Possibly a changing environment that shifts somewhat after a time (like year quarterly with daily schedule of tasks) to offer a variation of the costs of the different operations (movement and ?), but with a long enough interval for the 'learning'

The learning is the comparison of actual results with their picked solutions against similar 'runs' and trying to extract the most important factors.

Demonstration examples would be done on different 'maps' to show the feature as they get more complex (and also to simplify testing as you prove it works) - the pathfinding and planning via distances (time), then resources with different costs (now time and cost) then comparing/learning with uncertainty factor of supply - figure out which sources cause trouble by recording their history to develop an expectation pattern and then to have the AI consider that versatility pays out more regular than purely optimal sequences.

Then the adaption feature where the system rebuilds its knowledge against a changed/new map environment (where you have a discovery phase where you reevaluate(explore the new map) to then decide when you have enough info for the second phase of normal operation for the best results.