L’ Answer Set Programming (ASP) est un formalisme capable de représenter des connaissances en Intelligence Artificielle à l’aide d’un programme logique au premier ordre pouvant contenir des négations par défaut. En quelques années, plusieurs solveurs performants ont été proposés pour calculer les solutions d’un programme ASP que l’on nomme answer sets.Nous nous intéressons ici plus particulièrement au solveur ASPeRiX qui instancie les règles au premier ordre à la volée durant le calcul des answer sets. Pour réaliser cela, ASPeRiX applique un chaînage avant sur les règles à partir de littéraux précédemment déterminés.L’étude de ce solveur nous amène notamment à considérer la notion de justification dans le cadre d’une approche de calcul d’ answer sets basée sur les règles. Les justifications permettent d’expliquer pourquoi certaines propriétés sont vérifiées. Parmi celles-ci, nous nous concentrons particulièrement sur les raisons d’échecs qui justifient pourquoi certaines branches de l’arbre de recherche n’aboutissent pas à un answer set.Cela nous conduit à implémenter une version d’ ASPeRiX proposant du backjumping qui évite de parcourir systématiquement toutes les branches de l’arbre de recherche grâce aux informations fournies par les raisons d’échecs. / Answer set programming (ASP) is a formalism able to represent knowledge in Artificial Intelligence thanks to a first order logic program which can contain default negations. In recent years, several efficient solvers have been proposed to compute the solutions of an ASP program called answer sets. We are particularly interested in the ASPeRiX solver that instantiates the first order rules on the fly during the computation of answer sets. It applies a forward chaining of rules from literals previously determined. The study of this solver leads us to consider the concept of justification as part of a rule-based approach for computing answer sets. Justifications enable to explain why some properties are true or false. Among them, we focus particularly on the failure reasons which justify why some branches of the search tree does not result in an answer set. This encourages us to implement a version of ASPeRiX with backjumping in order to jump to the last choice point related to the failure in the search tree thanks to information provided by the failure reasons.

Much older soils data and soils information lies idle in libraries and archives and is largely unused, especially in developing countries like Kenya. We demonstrated the usefulness of a stepwise approach to bring legacy soils data ‘back to life’ using the 1980 <i>Reconnaissance Soil Map of the Busia Area</i> <i>(quarter degree sheet No. 101)</i> in western Kenya as an example. Three studies were conducted by using agronomic information, field observations, and laboratory data available in the published soil survey report as inputs to several digital soil mapping techniques. In the first study, the agronomic information in the survey report was interpreted to generate 10 land quality maps. The maps represented the ability of the land to perform specific agronomic functions. Nineteen crop suitability maps that were not previously available were also generated. In the second study, a dataset of 76 profile points mined from the survey report was used as input to three spatial prediction models for soil organic carbon (SOC) and texture. The three predictions models were (i) ordinary kriging, (ii) stepwise multiple linear regression, and (iii) the Soil Land Inference Model (SoLIM). Statistically, ordinary kriging performed better than SoLIM and stepwise multiple linear regression in predicting SOC (RMSE = 0.02), clay (RMSE = 0.32), and silt (RMSE = 0.10), whereas stepwise multiple linear regression performed better than SoLIM and ordinary kriging for predicting sand content (RSME = 0.11). Ordinary kriging had the narrowest 95% confidence interval while stepwise multiple linear regression had, the widest. From a pedological standpoint, SoLIM conformed better to the soil forming factors model than ordinary kriging and had a narrower confidence interval compared to stepwise multiple linear regression. In the third study, rules generated from the map legend and map unit descriptions were used to generate a soil class map. Information about soil distribution and parent material from the map unit polygon descriptions were combined with six terrain attributes, to generate a disaggregated fuzzy soil class map. The terrain attributes were multiresolution ridgetop flatness (MRRTF), multiresolution valley bottom flatness (MRVBF), topographic wetness index (TWI), topographic position index (TPI), planform curvature, and profile curvature. The final result was a soil class map with a spatial resolution of 30 m, an overall accuracy of 58% and a Kappa coefficient of 0.54. Motivated by the wealth of soil agronomic information generated by this study, we successfully tested the feasibility of delivering this information in rural western Kenya using the cell phone-based Soil Explorer app (<a href="https://soilexplorer.net/">https://soilexplorer.net/</a>). This study demonstrates that legacy soil data can play a critical role in providing sustainable solutions to some of the most pressing agronomic challenges currently facing Kenya and most African countries.<div><p></p></div>