Desenvolvimento de um ambiente virtual de tanques para treinamento de agentes inteligentes
In reinforcement learning, an agent is implemented with the aim of learning to perform some specified task in a given environment through the experiences obtained from interactions with that environment. The environment is the essential structure for this learning, since it is there that the funda...
Na minha lista:
Autor principal: | |
---|---|
Outros Autores: | |
Formato: | bachelorThesis |
Idioma: | pt_BR |
Publicado em: |
Universidade Federal do Rio Grande do Norte
|
Assuntos: | |
Endereço do item: | https://repositorio.ufrn.br/handle/123456789/53481 |
Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Resumo: | In reinforcement learning, an agent is implemented with the aim of learning to perform some specified task in a given environment through the experiences obtained from
interactions with that environment. The environment is the essential structure for this
learning, since it is there that the fundamental configurations for training agents are defined. One of these configurations is the choice of reward criteria and the definition of
action spaces. Considering a system of two coupled tanks as an environment and the
task specified as controlling the level of tank 1, training an agent in this real problem requires great care to avoid possible accidents in the laboratory. Some examples are level
overflow, incorrect voltages sent to the pump and possible loss of these tools. Thus, the
development of virtual environments is essential for training agents in this type of problem. With this, the objective of this work is to implement a virtual environment with the
Gymnasium (Gym) library of a system of coupled tanks to avoid possible accidents in the
laboratory and, with its graphical interface, facilitate the comparison of performance of
trained agents. For this, the identification of the tank system was used as a strategy for
modeling the system through two LSTM (Long-Short Term Memory) neural networks.
A network with only one LSTM layer for level prediction (single network) and another
network with an LSTM layer for each level (split network). Finally, the results obtained
from the training of the single and divided networks are presented, in addition to exposing the results of the Gym environment developed. It is also shown that the split network
served the purpose of modeling the tank system with a few millimeters error. |
---|