| Summary: | 碩士 === 長庚大學 === 資訊工程學系 === 98 === The major difficult of problem solving processes is the oversized search space corresponding to a complicated game state; therefore, efficient heuristics is always important to search algorithms. In this thesis, we conduct a series of novel experiments that train artificial neural networks to serve as heuristic functions. Adaptable heuristics is expected to reduce the states being retrieved and thus can make the problem solving process more effective.
Our experiments divide into two parts. The first part implemented Greedy Algorithms, and improved the heuristics thereof by consulting all the best children’s absolute positions and relative positions respectively. The basic idea is to move the states with worthless offspring backwords in the candidate queue; this way, we can avoid quantities of unnecessary retrievals on these states. The experimental results indicated that this method can reduce the searching nodes and time. However, the solutions obtained by greedy search were eventually much worst then the best solutions.
In the second part of our experiments, we conducted A* Algorithms.
Initially, artificial neural networks were successfully confirmed to simulate the currently known most effective heuristic function. Then we went a step futher to adjust the target values of the states that were involved in the near best solutions. Training like this enabled the near- best solutions to be found quickly. The experimental results revealed that the solutions obtained by this method are very close to the shortest path and both the retrieved nodes and search time were significantly reduced.
|
|---|
