Variable interaction is the dependence between optimization variables: changing one variable can alter another variable’s value or constraints. This dependence is also described as epistasis. section 18.1.10
House design
In a house-energy optimization problem, floor area and room count affect the amount of wood required for flooring. Changing the house size therefore forces a change in another variable when floor thickness stays fixed. section 18.1.10
Interaction changes both the optimization landscape and the choice of search method:
- Low interaction produces a smoother landscape; 0-1 knapsack is given as an example. section 18.1.10
- High interaction makes the landscape wavier and makes local exploration and fine-tuning harder; house design illustrates this case. section 18.1.10
- Use gradient descent for low epistasis, simulated annealing or random search for high epistasis, and genetic algorithms across medium-to-high epistasis. section 18.1.10
