Elysa Razafindrafara
Research Subjects
Elysa Razafindrafara has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Angelo Raherinirina
Centre de Recherche sur l’Enseignement des Mathématiques (CREM), University of Fianarantsoa, Fianarantsoa, Madagascar
4 shared publications - Mahefa Razafinirina4 shared publications
- Jean Berky Nguala
Laboratoire d’Informatique et de Mathématiques, Université de La Réunion, Réunion, FRANCE
3 shared publications Andrianiaina Ravintsoandraibe
1 shared publicationAngelo Raherinirina
1 shared publicationJean Berky Nguala
1 shared publication
Latest Publications
- Open Access
- 1 November 2026
- Published by Elsevier in Results in Control and Optimization
- Vol. 25
Trajectory-free identifiability and stability of expected hitting times in controlled absorbing Markov processes
- 1 November 2026
- Published by Elsevier in Results in Control and Optimization
- Vol. 25
Open AccessTrajectory-free identifiability and stability of expected hitting times in controlled absorbing Markov processes
- Open Access
- 4 June 2026
- Published by Elsevier in SSRN Electronic Journal
Do we need trajectories? Trajectory-free identifiability of expected hitting times in controlled absorbing Markov processes
Trajectory-level data is widely considered indispensable for recovering global dynamical quantities in stochastic control and reinforcement learning. This paper challenges that assumption by asking: are trajectories fundamentally necessary? We study controlled absorbing Markov processes under an extreme observational regime where each trajectory is reduced to a single transition, revealing only an initial state and its immediate successor, with no temporal links connecting observations. Despite this severe information constraint, we prove that the expected hitting time a central quantity in stochastic shortest-path problems remains identifiable at the population level from aggregated one-step observations under suitable coverage and persistence conditions. Our main contribution is a unified theoretical framework establishing four interlocking results: (i) identifiability of expected hitting times from first-order statistics alone; (ii) Lipschitz stability with explicit constants depending only on estimable quantities; (iii) structural bounds revealing how persistence and drift jointly control global dynamics; and (iv) a concentration phenomenon showing that hitting times are dominated by high-persistence states, uncovering a persistence bottleneck structure. The framework naturally extends to absorption probabilities and discounted value functions, suggesting a broader class of trajectory-free identifiable quantities.These results demonstrate that certain global dynamical quantities remain recoverable under substantially weaker temporal information than traditionally assumed, with implications for stochastic shortest-path problems, model-based reinforcement learning, and learning under severe observational constraints.
- 4 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDo we need trajectories? Trajectory-free identifiability of expected hitting times in controlled absorbing Markov processes
Trajectory-level data is widely considered indispensable for recovering global dynamical quantities in stochastic control and reinforcement learning. This paper challenges that assumption by asking: are trajectories fundamentally necessary? We study controlled absorbing Markov processes under an extreme observational regime where each trajectory is reduced to a single transition, revealing only an initial state and its immediate successor, with no temporal links connecting observations. Despite this severe information constraint, we prove that the expected hitting time a central quantity in stochastic shortest-path problems remains identifiable at the population level from aggregated one-step observations under suitable coverage and persistence conditions. Our main contribution is a unified theoretical framework establishing four interlocking results: (i) identifiability of expected hitting times from first-order statistics alone; (ii) Lipschitz stability with explicit constants depending only on estimable quantities; (iii) structural bounds revealing how persistence and drift jointly control global dynamics; and (iv) a concentration phenomenon showing that hitting times are dominated by high-persistence states, uncovering a persistence bottleneck structure. The framework naturally extends to absorption probabilities and discounted value functions, suggesting a broader class of trajectory-free identifiable quantities.These results demonstrate that certain global dynamical quantities remain recoverable under substantially weaker temporal information than traditionally assumed, with implications for stochastic shortest-path problems, model-based reinforcement learning, and learning under severe observational constraints.
- Open Access
- 27 May 2026
- Published by Elsevier in SSRN Electronic Journal
A Systematic Review of AI Usage for Educational Content in Low Resource Contexts: The Case of Madagascar
Background: Artificial intelligence (AI), particularly natural language processing (NLP) and deep learning, is increasingly used in education. However, its application to educational content generation in low-resource contexts remains insufficiently studied, especially in settings characterized by data scarcity, lowresource languages, and limited computational infrastructure, such as Madagascar. Objective: This systematic review analyzes existing research on AI-based educational content in low-resource contexts, with a focus on Madagascar, in order to identify dominant approaches, application domains, limitations, and research gaps under realworld constraints. Methods: Following PRISMA guidelines, studies published between 2015 and 2025 were retrieved from major scientific databases. including IEEE Xplore, ACM Digital Library, SpringerLink, Elsevier, and arXiv. Selected works focus on AIdriven educational content creation, adaptation, or personalization in low-resource environments. A qualitative synthesis was conducted to categorize methods, use cases, and contextual challenges. Results: The results indicate a predominance of NLP and deep learning methods, particularly transformer-based and large language models. Most studies target high-resource settings, with limited consideration of low-resource languages and infrastructures. Key challenges include data scarcity, high computational costs, and insufficient pedagogical evaluation. Lightweight models, prompt-based learning, black-box LLMs, and human-in-theloop strategies emerge as promising directions. Conclusions: This review reveals a clear gap in AI solutions tailored to low-resource educational contexts. It highlights the need for frugal, adaptable, and pedagogically grounded AI approaches, positioning Madagascar as a representative case for future research on inclusive and scalable AI-driven education.
- 27 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Systematic Review of AI Usage for Educational Content in Low Resource Contexts: The Case of Madagascar
Background: Artificial intelligence (AI), particularly natural language processing (NLP) and deep learning, is increasingly used in education. However, its application to educational content generation in low-resource contexts remains insufficiently studied, especially in settings characterized by data scarcity, lowresource languages, and limited computational infrastructure, such as Madagascar. Objective: This systematic review analyzes existing research on AI-based educational content in low-resource contexts, with a focus on Madagascar, in order to identify dominant approaches, application domains, limitations, and research gaps under realworld constraints. Methods: Following PRISMA guidelines, studies published between 2015 and 2025 were retrieved from major scientific databases. including IEEE Xplore, ACM Digital Library, SpringerLink, Elsevier, and arXiv. Selected works focus on AIdriven educational content creation, adaptation, or personalization in low-resource environments. A qualitative synthesis was conducted to categorize methods, use cases, and contextual challenges. Results: The results indicate a predominance of NLP and deep learning methods, particularly transformer-based and large language models. Most studies target high-resource settings, with limited consideration of low-resource languages and infrastructures. Key challenges include data scarcity, high computational costs, and insufficient pedagogical evaluation. Lightweight models, prompt-based learning, black-box LLMs, and human-in-theloop strategies emerge as promising directions. Conclusions: This review reveals a clear gap in AI solutions tailored to low-resource educational contexts. It highlights the need for frugal, adaptable, and pedagogically grounded AI approaches, positioning Madagascar as a representative case for future research on inclusive and scalable AI-driven education.
- Open Access
- 20 March 2026
- Published by Elsevier in SSRN Electronic Journal
Optimal policies and spectral convergence in finite absorbing MDPs: theory and educational applications
We present a comprehensive mathematical framework for adaptive education based on absorbing controlled Markov processes, motivated by the urgent need for instructional efficiency in resource-constrained environments such as Madagascar. Learning is modeled as a stochastic trajectory over a finite cognitive state space, driven by sequential pedagogical interventions interpreted as control actions. We formalize student progression as an absorbing Markov chain, deriving explicit expressions for the expected time to mastery via the fundamental matrix, and extend the model to a semi-Markov setting to capture the non-linear relationship between cognitive steps and instructional hours. Within this unified probabilistic framework, pedagogical decision-making is formulated as a finite Markov Decision Process (MDP). We establish: (i) almost-sure absorption toward mastery, (ii) the finiteness of expected learning time, and (iii) exponential convergence rates governed by the spectral radius ρ(Q). Empirical results from Reading and Mathematics remediation programs (TaRL) demonstrate that mathematical learning exhibits significantly higher "stochastic friction" (ρ ≈ 0.68), requiring nearly 40% more instructional time than literacy. We show that optimal policies minimize the spectral radius, effectively "linearizing" complex learning paths. Finally, we propose a didactic validation perspective to bridge spectral optimality with classroom-based pedagogical feasibility.
- 20 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessOptimal policies and spectral convergence in finite absorbing MDPs: theory and educational applications
We present a comprehensive mathematical framework for adaptive education based on absorbing controlled Markov processes, motivated by the urgent need for instructional efficiency in resource-constrained environments such as Madagascar. Learning is modeled as a stochastic trajectory over a finite cognitive state space, driven by sequential pedagogical interventions interpreted as control actions. We formalize student progression as an absorbing Markov chain, deriving explicit expressions for the expected time to mastery via the fundamental matrix, and extend the model to a semi-Markov setting to capture the non-linear relationship between cognitive steps and instructional hours. Within this unified probabilistic framework, pedagogical decision-making is formulated as a finite Markov Decision Process (MDP). We establish: (i) almost-sure absorption toward mastery, (ii) the finiteness of expected learning time, and (iii) exponential convergence rates governed by the spectral radius ρ(Q). Empirical results from Reading and Mathematics remediation programs (TaRL) demonstrate that mathematical learning exhibits significantly higher "stochastic friction" (ρ ≈ 0.68), requiring nearly 40% more instructional time than literacy. We show that optimal policies minimize the spectral radius, effectively "linearizing" complex learning paths. Finally, we propose a didactic validation perspective to bridge spectral optimality with classroom-based pedagogical feasibility.