Abstract
This work presents some approaches to overcome current Reinforcement Learning limits. We implement a simple virtual environment and some state-of-the-art Reinforcement Learning algorithms for testing and producing a baseline for comparison. Then we implement a Relational Reinforcement Learning algorithm that shows superior performance to the baseline but requires introducing human knowledge. We also propose that Model-based Reinforcement Learning can help us overcome some of the barriers. For better World models, we explore Inductive Logic Programming methods, such as First-Order Inductive Learner, and develop an improved version of it, more adequate to Reinforcement Learning environments. Finally we develop a novel Neural Network architecture, the Inductive Logic Neural Network, to fill the gaps of the previous implementations, that shows great promise. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020.

Abstract
The present work surveys research that integrates successfully a number of complementary fields in Artificial Intelligence. Starting from integrations in Reinforcement Learning: Deep Reinforcement Learning and Relational Reinforcement Learning, we then present Neural-Symbolic Learning and Reasoning since it is applied to Deep Reinforcement Learning. Finally, we present integrations in Deep Reinforcement Learning, such as, Relational Deep Reinforcement Learning. We propose that this road is breaking through barriers in Reinforcement Learning and making us closer to Artificial General Intelligence, and we share views about the current challenges to get us further towards this goal. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020.

Abstract
Types in logic programming have focused on conservative approximations of program semantics by regular types, on one hand, and on type systems based on a prescriptive semantics defined for typed programs, on the other. In this paper, we define a new semantics for logic programming, where programs evaluate to true, false, and to a new semantic value called wrong, corresponding to a run-time type error. We then have a type language with a separated semantics of types. Finally, we define a type system for logic programming and prove that it is semantically sound with respect to a semantic relation between programs and types where, if a program has a type, then its semantics is not wrong. Our work follows Milner’s approach for typed functional languages where the semantics of programs is independent from the semantic of types, and the type system is proved to be sound with respect to a relation between both semantics.

Abstract
There is a growing need for fast and accurate methods for testing developmental neurotoxicity across several chemical exposure sources. Current approaches, such as in vivo animal studies, and assays of animal and human primary cell cultures, suffer from challenges related to time, cost, and applicability to human physiology. Prior work has demonstrated success employing machine learning to predict developmental neurotoxicity using gene expression data collected from human 3D tissue models exposed to various compounds. The 3D model is biologically similar to developing neural structures, but its complexity necessitates extensive expertise and effort to employ. By instead focusing solely on constructing an assay of developmental neurotoxicity, we propose that a simpler 2D tissue model may prove sufficient. We thus compare the accuracy of predictive models trained on data from a 2D tissue model with those trained on data from a 3D tissue model, and find the 2D model to be substantially more accurate. Furthermore, we find the 2D model to be more robust under stringent gene set selection, whereas the 3D model suffers substantial accuracy degradation. While both approaches have advantages and disadvantages, we propose that our described 2D approach could be a valuable tool for decision makers when prioritizing neurotoxicity screening. © 2019 IEEE.

Abstract
Diabetes type I is a chronic disease that requires strict supervision. MyDiabetes is a utility application for diabetic users. This application served as basis to develop a logical unit, composed of logical rules, translated from medical protocols and guidelines, to advise the user. The data in the application is a source of knowledge about the user's health state and diabetes intrinsic characteristics. An existing concern is the weak user adherence and consequential data absence. The implemented solutions were gamification and an interface rework. As later confirmed through a survey, users feel captivated by appealing interfaces, achievements and medals. In a near future, we will resume our work with the S. Joao's hospital, with a new trial and volunteers. User testing will be used to validate the gamification techniques.

Abstract
The increasing use of sensor networks has led to an ever larger number of available spatiotemporal datasets. Forecasting applications using this type of data are frequently motivated by important domains such as environmental monitoring. Being able to properly assess the performance of different forecasting approaches is fundamental to achieve progress. However, traditional performance estimation procedures, such as cross-validation, face challenges due to the implicit dependence between observations in spatiotemporal datasets. In this paper, we empirically compare several variants of cross-validation (CV) and out-of-sample (OOS) performance estimation procedures, using both artificially generated and real-world spatiotemporal datasets. Our results show both CV and OOS reporting useful estimates, but they suggest that blocking data in space and/or in time may be useful in mitigating CV's bias to underestimate error. Overall, our study shows the importance of considering data dependencies when estimating the performance of spatiotemporal forecasting models.