Searching A Solution Space
Published: 07 Jul 2013
In chapter five of Programming Collective Intelligence I was introduced to new ways to go about solving optimization problems. When I say optimization problems what I’m really saying is that there is problems where there is a function to be minimized with respect to a large number of possible solutions. The knapsack problem would be one example of an optimization problem.
Sometimes the search space is so vast that checking every solution would take too long. That is where a stochastic optimization comes in. Perhaps the simplest example of a stochastic optimization is taking a simple random sample of the solution space and then checking each of the samples to see how much they would cost and returning the one that was best. The key idea being that reducing the search space for the sake of saving computation at the cost of accuracy.
There are much cleverer stochastic optimizations then random searching which take advantage of the structure of the search space. For some problems, if you were to picture the search space and then add another dimension to it which corresponded with the fitness of the solution the result could be something akin to a landscape with hills and valleys. The hills would be where the cost function was low and the valleys where the cost function was high.
For example, take hill climbing. In hill climbing, starting from a random point in the solution space, you look at the surrounding solutions in the solution space instead of all the solutions and try to head towards the highest-quality solutions in an iterative process. This approach is still stochastic, because you start from a random point. By moving down the hills into valleys the hill climbing algorithm will always be able to reach a local minimum where the solution is at least as good as the solution of it’s neighbors.
This algorithm would be like dropping a ball onto the geography. It would roll downward until coming to rest.
This algorithm, as one can imagine, will not always reach the lowest point. It can get stuck in a local optima. There are other stochastic optimization algorithms which try to reduce the likelihood of getting stuck in local optima. Included in this set of algorithms are simulated annealing and genetic algorithms.