Reinforcement Learning enables to train an agent via interaction with the environment. However, in the majority of real-world scenarios, the extrinsic feedback is sparse or not sufficient, thus intrinsic reward formulations are needed to successfully train the agent. This work investigates and extends the paradigm of curiosity-driven exploration. Our aim is to develop means for the better incorporation of state-and/or action-dependent information into existing intrinsic reward formulations. First, a probabilistic approach is taken to exploit the advantages of the attention mechanism, which is successfully applied in other domains of Deep Learning. Combining them, we propose new methods, such as Attention-aided Advantage Actor-Critic, an extension of the Actor-Critic framework. Second, another curiosity-based approach - Intrinsic Curiosity Module - is extended. The proposed model utilizes attention to emphasize features for the dynamic models within Intrinsic Curiosity Module, moreover, we also modify the loss function, resulting in a new curiosity formulation, which we call rational curiosity (RCM). The corresponding implementation can be found at https://github.com/rpatrik96/AttA2C/.

Driver fatigue reduces the safety of traditional driving and limits the widespread adoption of self-driving cars; hence, the monitoring and early detection of drivers’ drowsiness plays a key role in driving automation. When representing the drowsiness indicators as large feature vectors, fitting a machine learning model to the problem becomes challenging, and the problem’s perspicuity decreases, making dimensionality reduction crucial in practice. For this reason, we propose an embedded feature selection algorithm that can be later utilized as a building block in the system development of a neural network-based drowsiness detector. We have adopted a technique: a so-called Feature Prune Layer is placed in front of the first layer in the architecture; as a result, its weights change regarding the importance of the corresponding input features and are deleted iteratively until the desired number is reached. We test the algorithm on EEG data, as it is one of the best indicators of drowsiness based on the literature. The proposed FS algorithm is able to reduce the original feature set by 95% with only 1% degradation in precision, while the precision increases by 1.5% and 2.7% respectively when selecting the top 10% and top 20% of the initial features. Moreover, the proposed method outperforms the widely popular Principal Component Analysis and the Chi-squared test when reducing the original feature set by 95%: it achieves 24.3% and 3.2% higher precision respectively.

RoboCup is one of the major global AI events, gathering hundreds of teams from the world’s best universities to compete in various tasks ranging from soccer to home assistance and rescue. The commonality of these three seemingly dissimilar tasks is that in order to perform well, the robot needs to excel at the all major AI tasks: perception, control, navigation, strategy and planning. In this work, we focus on the first of these by presenting what is—to our knowledge—the first fully neural vision system for the Nao robot soccer. This is a challenging task, mainly due to the limited computational capabilities of the Nao robot. In this paper, we propose two novel neural network architectures for semantic segmentation and object detection that ensure low-cost inference, while improving accuracy by exploiting the properties of the environment. These models use synthetic transfer learning to be able to learn from a low number of hand-labeled images. The experiments show that our models outperform state-of-the-art methods such as Tiny YOLO at a fraction of the cost.