Improving tower defense game AI (Genetic algorithm vs Genetic programming)

Tio, Chun Chieng (2015) Improving tower defense game AI (Genetic algorithm vs Genetic programming). Universiti Malaysia Sabah. (Unpublished)


Download (913kB) | Preview


The processes of designing and developing in digital game is costly. Tower defense games received much attention recently. However, there are few challenges found in designing the maps used in the game; the maps used is either too easy to be played or too difficult to be won. It happened as programmers simply developed the maps without proper planning and testing. This research proposed a technique using Evolutionary Algorithms (EAs) and Artificial Neural Networks (ANNs) to auto generate controllers to test the proposed maps. The proposed method can lead to significantly better intelligent system rather than depending on either EAs or ANNs alone. The selected EAs are Genetic Algorithm (GA) and Genetic Programming (GP) and the selected ANNs are Feed-Forward Neural Network (FFNN), Elman Recurrent Neural Network (ERNN), Jordan Recurrent Neural Network (JRNN), and an Ensemble Neural Network (ENN). The proposed ENN is a weighted sum NN composed of single FFNN, single ERNN, and single JRNN. The elitism concept is integrated in the optimization processes. The GA utilized uniform mutation and uniform crossover operators whereas the GP used one point mutation and one point crossover operators. Each experiment is conducted 10 times for each algorithm. The boxplot and t-test are used as performance metric in this project. As a result, by comparing the proposed EAs with same ANN used, the results showed GA performed slightly better than GP. On the other hand, by comparing different ANNs in the GA experiment, GA hybrid ENN achieved 91% of success rate, GA hybrid FFNN achieved 89% of success rate, GA hybrid JRNN achieved 81% of success rate, and GA hybrid ERNN achieved 70% of success rate. Hence, it concludes that GA hybrid ENN outperformed other type of NNs. Interestingly, the GP experiments showed different results. The GP hybrid FFNN achieved highest average success rate which is 88%, GP hybrid ENN achieved 67% success rate, GP hybrid ERNN achieved 64% success rate, and GP hybrid JRNN achieved 63% success rate. Overall comparison showed GA hybrid ENN outperformed the GP hybrid FFNN.

Item Type: Academic Exercise
Uncontrolled Keywords: intelligent system, Genetic Algorithm, Genetic Programming, elitism concept, hybrid
Subjects: Q Science > QA Mathematics > QA76 Computer software
Divisions: FACULTY > Faculty of Computing and Informatics
Depositing User: Unnamed user with email
Date Deposited: 05 Nov 2015 06:56
Last Modified: 27 Oct 2017 09:05

Actions (login required)

View Item View Item