Artificial Intelligence usually has to sort through a great many options for a good solution. We may have to go through all those options, which is very time-intensive; instead, AI uses heuristics in search, which in turn produce results faster. One such heuristic is Hill Climbing, which is a very simple yet very powerful local search algorithm.
The idea is easy to understand: Imagine at a point on a mountain peak which you are standing at and instead of looking at the whole landscape you look at the nearby paths and you keep moving in that which is up. In AI, this is what hill climbing does, which is to continuously pick a better neighbouring solution.
What Is Hill Climbing in AI?
Hill Climbing, which is a heuristic local search algorithm, improves the present solution at each step by looking at a neighbouring state. We look at what is available nearby and choose the one that does better according to the objective or evaluation function.
The present algorithm is oriented towards the current solution, which in turn means we see speed and memory efficiency, but also that it may not always find the best possible solution.
For instance, in the case of AI looking for the shortest path through many cities, it may start with a given route, which it then changes slightly by, for example, switching the order of two cities out of the whole route, and will keep that change if it in fact reduces the total distance.
How Does the Hill Climbing Algorithm Work?
The process breaks down into a few simple steps:
- Choose an initial state: At random, the algorithm will select a solution.
- Evaluate the state: A metric that evaluates the present solution.
- Generate neighbouring states: Tiny shifts are introduced, which in turn present other options.
- Compare the solutions: The algorithm looks at the adjacent states.
- Move to a better state: When an improvement is found, it becomes the current state.
- Stop when necessary: At which point the process stops when no better solution is found, or we hit another stop condition.
This is a greedy approach which at each step focuses on immediate improvement instead of all possible future options.
Types of Hill Climbing
Simulated Hill Climbing.
The algorithm goes through each neighbouring state in order and transitions to the first that improves the current solution the most. It is easy to implement but may not always find the best improvement.
Steepest Ascent Hill Climbing.
In this case, we look at all relevant neighbouring states, and the algorithm chooses the one that has the greatest improvement. We do produce better local solutions but at a higher computational cost.
Random Hill Climbing.
Instead of always going with the best option, we introduce an element of chance in which we choose which improved state to go to. This can put more variety into the search.
Random Reboot Hill Climbing.
The algorithm we run is hill climbing, which we do many times from various start points. If one go at it gets stuck in a poor solution, another start point may present a better one.
Common Problems in Hill Climbing
Although very simple, hill climbing has a number of issues.
- Local Maximum: The algorithm may get to a state which is an improvement over its nearby solutions but which isn’t the best we have.
- Plateau: At times we see that adjacent states have equal evaluation. When there is no clear improvement, the algorithm will get stuck.
- Ridge: In some cases we see that better solutions present themselves in a small area, but the put forth moves may not allow the algorithm to easily follow that in.
- Dependence on Initial State: Different results from different start points. A poor initial solution may result in a poor final solution.
These issues are what make hill climbing a non-choice for finding the global optimum.
Applications of Hill Climbing in AI
Hill climbing is applied to a wide range of optimisation problems, which include route planning, scheduling, robotics, game-related optimisation, feature selection, and machine learning hyperparameter tuning.
For example, in the case of scheduling, we may change an initial timetable through the use of reordering or trade-off of tasks. The algorithm will keep any changes that improve conflict resolution or resource use.
How Can Its Limitations Be Reduced?
Techniques like that of random restarts, which allow for limited lateral moves, stochastic selection and simulated annealing improve exploration. Simulated annealing in particular is a good choice as at times it will accept a worse solution, which in turn allows the search to leave a local maximum.
Conclusion
Hill Climbing is a practical approach which allows AI systems to improve solutions without the need to search the full space. It is greedy in nature, which makes it simple and efficient; also, its low memory requirement is a great attribute for many optimisation tasks. But we see that issues such as local maxima, plateaus, ridges, and the start point issue do in fact affect results.
Make a little improvement at a time, evaluate the result, and keep going until progress ceases. That may not always identify the best possible solution, but with different variations and in combination with other algorithms, it can be very useful as an element of practical AI optimisation systems.
About the Author
