Aplicação de metodos de computação flexivel em navegação autonoma de veiculos

AUTOR(ES)
DATA DE PUBLICAÇÃO

1995

RESUMO

Autonomous vehicle navigation is a well-known and typical example of an autonomous control problem. This type of control problem involves a very complex, unstructured environment. The great amount of parameters preclude the use of a mathematical model of the environment. Thus, autonomous control demands constantly adapting, locally situated control methods, giving rise to two fundamental design issues: situatedness and frame-of-knowledge. The first deals with the impossibility of predicting the global behaviour of the environment. Instead the autonomous agent must be able to act upon local and limited data. The latter problem refers to the knowledge used in the design of an agent (a priori knowledge). Such pre-installed experience limits the agent s adaptability by imposing the designer s point of view (based on sensors and processing modes far different from those of the agent) and limited knowledge of the environment. Both issues are addressed by implementing an adaptive locally-aware control method. Qne of the most promising approaches comes from the 80ft computing field. Soft computing involves a spectrum of methods, ranging from fuzzy system theory to evolutionary systems. The main bulwarks are neural networks, fuzzy systems and a number of methods known as probabilistic reasoning. The latter includes evolutionary systems, cellular automata, complexity theory and chaos theory, among others. The power and promise of soft computing emerge from the symbiosis of its many paradigms. These hybrid synergetic systems offer the strengths of their components while cross-compensating the components drawbacks. The most interesting features include auto-organization, imprecise data handling and unsupervised learning capabilities. In this work we propose an autonomous control method that combines a neural- network, fuzzy system theory and a genetical algorithm. The synergy of these three soft computing paradigms offers a parallel robust computing structure with easily extractable/insertable knowledge and capable of unsupervised learning. The proposed method was applied to the autonomous vehicle navigation problem in a simulated environment. The results show that most of the limitations are due to the amount and type of knowledge used in the choice of the vehicle parameters (sensors and knowledge representation). The navigation method itself proved to be robust and reliabJe, deveJoping behaviours comparable to those of a hand-crafted agent. Future work will focus on the issues of the optimum a priori knowledge levels, the sensivity of the learning method to sensor specifications, and on the development of self-organizing methods for relevant environment data recognition. These efforts are geared towards freeing the agent from human-imposed limiting factors while enhancing its adaptability

ASSUNTO(S)

automação computação sistemas de veiculos auto-guiados redes neurais (computação) detectores

ACESSO AO ARTIGO

Documentos Relacionados