The Problem

Your unit has a limited training budget. You can resource 3 events on your own, or 4 if you call in a favor from a sister company. That 4th event costs extra overhead. Maximize total readiness by choosing the right events and allocating hours wisely.

  • Each event has diminishing returns . More hours help, but with less and less impact.
  • Some event combinations have hidden synergies that boost overall readiness.
  • Events you don't select suffer atrophy . Skills degrade without practice.
  • Selecting a 4th event requires calling in a favor . This costs 5-15 points of overhead.
  • Every evaluation is noisy . The same allocation can produce different scores each time.
Individual Payoff Curves

These are what you CAN observe in isolation.

Your Simulation Results

Each dot is a single training quarter. The GA evaluates fitness as the mean across N quarters, which is much less noisy.

What is a Chromosome?

A genetic algorithm represents each candidate solution as a chromosome . Each chromosome has 10 genes:

  • Genes 1-5: binary (0 = no, 1 = yes). 3 or 4 must be selected (4 incurs overhead).
  • Genes 6-10: hours allocated to each event. Must sum to the budget.
  • Unselected events have weight genes but they are zeroed out in decoding.
Example Chromosome
How Selection Works

Selection decides which individuals get to be parents for the next generation. Better solutions should be chosen more often, but not always. Some randomness preserves diversity and prevents premature convergence.

Population Fitness
Genetic Operators
How Crossover Works

Crossover combines two parents to produce one child .

Parent 1 + Parent 2 = Child
Decoded Allocations
How Mutation Works
Before and After
Where Did the Draws Land?
Fitness Impact
Convergence
Best Allocation So Far
Population Diversity

Drops as the population converges.

Current Generation's Best

Updates every generation as new candidates are explored.

What to Look For

Each config runs independently on the same problem. Compare:

  • How fast does each config approach the optimum?
  • Does it plateau early or keep improving?
Convergence Comparison
Results Summary