Go home

Local Search

Overview

• Search through neighboring configurations for an improvement to an objective function
• Only keep track of a single search state (no history)

Hill Climbing

• Start at a random state
• Move to a neighboring configuration that improves the objective function
• Discard a step that results in a worse configuration

${n}$-Queens

• Start with random configuration of one queen in each column and row
• Swap queens' rows, keeping track of number of diagonal conflicts

Linear regression

• Start with a horizontal line
• Adjust slope and ${y}$-intercept until the ${R^2}$ value is minimized

Traveling Salesman

• Start with a random configuration (list of nodes to traverse)
• Swap pairs of nodes if it results in a beneficial configuration

Pros and Cons of Local Search

Pros:

• Constant space
• Can easily be tweaked with new data

Cons:

• Can get stuck in local optima (where all neighbors look worse)
• Can get stuck in plateaus (where all neighbors look the same)

Local Optima

Methods to get out of local optima:

• Make a large change when small changes stop working
• Allow some sequence of bad steps once in a while (but keep a copy of the best solution)
• Restart the search (either from another initial condition or the last best search)

Simulated Annealing

Idea comes from metal: as metal cools, it becomes harder for molecules to move around as the temperature decreases, resolving internal stresses gradually.

Allow moving to a worse state with a small probability, but have this probability decrease as the process continues.

1. Initialize a temperature ${T}$ and a decay multiplier ${d}$
2. Allow moving to a worse state with probability ${e^{{e(n)-e(n^\prime)}/{T}}}$ where the ${e}$ in the exponent is the energy function (objective function to minimize)
3. Probability decays exponentially as the process continues

Local Beam Search

Similar to regular beam search, except keep the best ${k}$ options across all explored branches. This is essentially running ${k}$ searches in parallel, but the searches can communicate with one another.

Genetic Algorithms

Just like evolution in biology, keeps the best configurations and combines them with random mutations to achieve a better solution.

• Fitness: the objective function to maximize
• Mutation: changing to a neighboring solution at random
• Crossover: combining two configurations to produce a new configuration with some parts of each parent