| Full text | |
| Author(s): |
Silva, Isaac J.
;
Vilao Junior, Claudio O.
;
Costa, Anna H. Reali
;
Bianchi, Reinaldo A. C.
Total Authors: 4
|
| Document type: | Journal article |
| Source: | JOURNAL OF CONTROL AUTOMATION AND ELECTRICAL SYSTEMS; v. 32, n. 4, p. 11-pg., 2021-04-16. |
| Abstract | |
One of the challenges of Deep Learning research is to develop algorithms for mobile robotic agents that operate in uncontrolled environments, in which dynamic changes and limited processing power are common restrictions. A common solution is to develop separate vision and decision modules, so that the former is based on deep neural network architectures and the latter is based on rules, and then interconnect them. The drawback of this solution is that the modules need to exchange high-level information about the objects in the scene, which are usually the positions of all objects in the scene, and this is computationally expensive. To address this problem, this paper presents a Decision Tree of Deep Neural Networks (DT-DNN) that aims to perform end to end-from image to decision-processing, and, thus, eliminating the need for quantitative and relational information about the image. This model is composed of smaller and more specialized modular DNNs, thus solving the trade-off between performance and inference time. Experiments were carried out using a real robot in the RoboCup Humanoid League domain in a soccer field, and also in simulation. We compared DT-DNN with several traditional DNN architectures. From the results, it is possible to conclude that the use of the DT-DNN made the system simpler and more robust, with fewer parameters to be adjusted, reducing the time spent with inference and also increasing the performance when compared to the traditional approach. (AU) | |
| FAPESP's process: | 19/07665-4 - Center for Artificial Intelligence |
| Grantee: | Fabio Gagliardi Cozman |
| Support Opportunities: | Research Grants - Research Program in eScience and Data Science - Research Centers in Engineering Program |