A genetic‐fuzzy system for optimising agent steering

Abstract

Fuzzy controllers are a popular method for animated character perception and control. A drawback of fuzzy systems is that time intensive manual calibration of system parameters is required. We introduce a new Genetic-Fuzzy System (GFS) that automates the tuning process of rules for animated character steering. An advantage of our GFS is that it is able to adapt the rules for steering behaviour during run time. We explore the parameter space of the new GFS, and discuss how the GFS can be implemented to run as a background process during normal execution of a simulation. Copyright © 2010 John Wiley & Sons, Ltd.KEY WORDS: agent navigation and steering; genetic algorithms; fuzzy logic IntroductionFuzzy logic controllers are an elegant solution for agents controlling the motion of real time animated characters, creatures and vehicles because fuzzy controllers produce outputs that transition smoothly, rather than stepping between output states, and because they can be implemented to allow agents to react to changing environments and moving obstacles in real time, with minimal processing overheads 1 . This approach is ideal for modern games and 3D simulations that need to simulate large crowds of animated characters, where each character requires its own realistic motion 2 . An outstanding problem with the use of fuzzy controllers for steering and moving animated characters is that the controllers need to be tuned to suit each new type of agent's combination of rôle, physical and performance characteristics, and operating environment. This means that, while the essential kernel of the system-the fuzzy decision making process-applies broadly, all the parameters of the fuzzy systems need to be tailored to suit the peculiarities of each new type of agent.Building and calibrating fuzzy controllers consists of the following steps 3 :(1) Designing the size and scale of fuzzy set functions that classify ranges of real (crisp) values, such as dis- This process is very time consuming for the designer. Modifying a system by trial and error based on test cases has taken 5-10 hours in our applications. If there are various types of agent involved in a simulation, or an agent has to cope with a large variety of different cases then the required tuning time multiplies. There is no guarantee that hand tuned rules and sets are optimal. The fuzzy controllers are also not able to adapt to any change to the environment after tuning.Our specific aims in this work are:(1) To automate the tuning process of fuzzy controllers used for steering and moving agents in a 3D simulation, saving developer time and improving manually calibrated controllers. (2) To design an architecture that allows this tuning process to happen in real time, giving the agents an adaptive quality. A potential solution to the fuzzy tuning problem could be to use an evolutionary algorithm to automatically tune the fuzzy controllers. The novel contribution of this work is that it provides an automatic calibration mechanism for fuzzy controllers specific to 3D animated agen...

Crowd simulation for crisis management: The outcomes of the last decade

“…A Genetic-Fuzzy System for agent steering, that automated the tuning process of the rules was introduced in 2010, (Gerdelan & O'Sullivan, 2010). The simulation results showed that such setup was very useful and efficient for reducing the human interface with the simulation environment for further tuning.…”