Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/85468
Title: Game strategy indexing, learning and optimization in real time strategy (RTS) games using soft computing techniques
Authors: Ng, Hiu Fung
Degree: Ph.D.
Issue Date: 2013
Abstract: In real-time strategy (RTS) games, players position and maneuver units and structures under their control to secure areas and destroy their opponents' assets. Typical strategies are resource gathering, units formation and positioning, base building, technology development and path finding. All the movements, construction and researches take place in real time, and players have a bird's eye view to control and monitor units using their own strategy. Selecting which strategy to use becomes one of the major challenges. This research focuses on indexing, learning and optimization of these RTS games strategies using soft computing techniques. Three game strategies are selected for the investigation, and original techniques have been developed to tackle this strategy determination problem. Based on our findings, development of RTS computer game software is better understood and supported using soft computing techniques. The first investigation is to develop a strategy to quickly position game units effectively in a map so that they will create maximum casualty to enemies. A model integrating artificial neural network (ANN), genetic algorithm (GA) and case-based reasoning (CBR) is proposed and tested. The main idea is to evaluate the past strategies using GA, and train up an ANN for fast retrieval of units' locations. When new maps and new conditions are presented, CBR is used to compute the adjustments needed for the new locations. The key contribution here is the formulation of the RTS game strategy selection as CBR planning using a neural-evolutionary model. A number of simulated experiments with different maps and game unit settings are carried out to test the model. The result demonstrated that the model provides an efficient and natural game strategy indexing and determination scheme.
The second investigation is to develop a strategy to determine the types of game units to be selected for production with the purpose to effectively combat with opponents' troops. A contribution is made here by considering how the order of production and feature interaction of game units affect the result of playing RTS games. Due to complicated game rules, extensive terrains and numerous playable items, exhaustive search or explicit description of unit combination effects using analytical models, such as finite state machines, Bayesian networks and decision trees may not be feasible. We developed a machine learning model that extracts and evaluates game unit combination strategy from past data. This model takes into account the sequence in which game units are produced and the interaction among them. We combine fuzzy measure, fuzzy integral and genetic algorithm to develop the model. Warcraft III battle data from real players are used in our experiments. Compared with the traditional Choquet Integral, our new order-based fuzzy integral gives a smaller training and testing error in RTS game strategy selection. A dynamic Bayesian Network is also developed in learning game players' behavior. The third investigation is optimal path determination. This is complicated because RTS game environment is hostile, dynamic and consists of many different types of game units interacting with each other in the battle field. Traditional path searching algorithms like min-max, alpha-beta pruning, hill climbing and A* are not suitable in such a complicated dynamic game world. We modified the multi-agent potential field model by incorporating the non-linear feature interaction property. The effect of unit cooperation can then be described, and therefore taken into consideration in optimal path determination. Our approach can identify the direction of positive and negative interaction for unit movement planning and team composition in RTS games. A combination of using real data and simulation experimental setting is used in this investigation. The results demonstrated that our path determination method is much better than the traditional methods implemented in Warcraft III. As a summary, this PhD research focused on the investigation of RTS game strategies. Three original models are developed, namely (i) a neural-evolutionary model for CBR planning, (ii) an order-based fuzzy integral model, and (iii) a model of multi-agent potential field with feature interaction. All these three models are tested experimentally and promising results were obtained. A number of conference papers were published. One journal paper is under second review while another one is under preparation.
Subjects: Computer games -- Programming.
Hong Kong Polytechnic University -- Dissertations
Pages: 178 leaves : ill. ; 30 cm.
Appears in Collections:Thesis

Show full item record

Page views

50
Last Week
1
Last month
Citations as of Mar 24, 2024

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.