Integrating neural networks into the agent’s decision-making: A Systematic Literature Mapping
Rodrigo Machado Franco Rodrigues, Ricardo Azambuja Silveira, Rafael de Santiago · 2021
AI systems have been playing a crucial role in many different fields of study. Even though connectionist methods, more precisely deep neural networks, are more prevalent nowadays, many of their limitations have delayed the deployment of AI systems in relevant areas, such as healthcare, financial, and legal. One of its main criticisms relies on the fact that deep neural networks require large data sets, poor generalization, and lack of interpretability. Researchers believe that the next level of AI will require integrating these connectionist methods with different AI’s fields. Although many different studies explore this research topic, many of them are surveys or do not cover AI’s new advances. A Systematic Literature Mapping is performed to fill this gap, which aims to explore the integration of neural networks into the intelligent agent’s decision making. In this study, we analyzed over 1000 papers, and the main findings are: (i) 64% of studies use neural networks to define the learning agent’s reward policies; (ii) 5% of studies explore the integration of neural networks as part of the agent’s reasoning cycle; and (iii) although 55% of studies main contributions are related to neural networks and agents design, we find that the remaining 45% of the studies use both agents and neural networks to solve or contribute to a particular field of study or application.