SynthEd: From Synthetic Data to Simulated Learners

Read full text Summarise with VIRAT
Authors: h_aykut_cosgun, Evrim Genç Kumtepe
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

SynthEd is an agent-based simulation environment for open and distance learning (ODL) research. By combining persona-driven agent modeling with 10 established theoretical frameworks, it generates behaviorally grounded and temporally coherent learning trajectories. Each synthetic student's engagement, dropout risk, GPA, and social interactions emerge from evolving motivations, experiences, and life contexts. Simulation Engine 10 theory modules (Tinto, Bean & Metzner, Kember, SDT, Garrison CoI, Moore, Rovai, Bäulke, Epstein & Axtell, Gonzalez) with 4-phase protocol dispatch and auto-discovery Continuous persona spectrum: employment intensity, family responsibility, internet reliability as [0,1] floats with Beta distributions Multi-semester simulation with carry-over mechanics for engagement, GPA, coping, dropout phases GPA feedback loop anchoring cost-benefit, non-fit perception, and competence beliefs Calibration & Validation Sobol sensitivity analysis (68 parameters) + NSGA-II multi-objective optimization with hypervolume convergence tracking; Pareto front re-evaluated at N=2000 across 3 seeds for noise-free knee-point selection; replicated calibration with cross-seed comparison; 10-seed validation run 5-level validation suite: 22 statistical tests (distributions, correlations, temporal coherence, privacy, backstory) Adaptive parameter bounds adjusting calibration space to demographic changes Configuration & Output InstitutionalConfig (5 quality params), GradingConfig (Beta/Normal/Uniform, dual-hurdle, relative grading), EngineConfig (70 constants), PipelineConfig (JSON serialization) OULAD-compatible 7-table CSV export, optional LLM-enriched backstories, benchmark reports Interactive dashboard (Shiny for Python + Plotly): dark-themed UI with configuration presets, reactive distribution validation, 4 summary cards and charts, JSON config import/export v1.7.0 Highlights Calibration methodology document gains a new §7 "Limitations & Identifiability" section: explicitly documents the 18-D parameter null space (20 free parameters fit to 2 scalar objectives), the simulator's empirical Monte Carlo noise floor (≈1.5–2 pp dropout σ at n=1000), and the resulting parameter-interpretation caveats for downstream users Aggregation bug fix: combined-profile JSON now contains all calibration seeds with explicit seed identifier (previously only the last seed was retained) Cross-seed knee-point distance demoted from a binary release gate to an informational diagnostic, reflecting that cross-seed parameter scatter is a structural property of the model rather than an optimizer regression Dashboard fixes: light-theme navbar contrast (1.1:1 → 17:1+), tablet 768 px run-bar overlap resolved, secondary-text contrast raised across all theme variables Dashboard mode-split skeleton: "Configure" tab renamed to "Research", new "Calibrate" placeholder tab as scaffolding for upcoming OULAD-indexed calibration tooling v1.8.0 Highlights Auto-generated test inventory table in THEORY.md via doc_facts — sentinel-demarcated, AST-extracted module docstrings, eliminates manual table drift SSQ integration test thresholds empirically calibrated with a 50-seed study: eps lowered from 0.010 to 0.008 with >90% statistical power Calibrate tab validation scorecard (PR B): collapsible accordion table with pass/fail markers, multiple-testing footer, empty-state handling Dashboard accessibility pack: WCAG 2.4.7 focus-visible, tooltip ARIA, numeric-input locale fix, section heading contrast analysis/ coverage uplift: nsga2_calibrator 69%↔94%, sobol_sensitivity 83%↔98% 879 pytest tests across 49 files v1.8.1 Highlights Revived a silently-skipped validation test: the engagement–GPA correlation now actually runs (the pipeline emits the field the validator reads) Hardened Sobol failed-sample handling — failed, crashed, or timed-out samples are retried in isolation on a fresh executor instead of zero-filling (which contaminated the S1/ST indices) or aborting on one worker death; fail-fast only when genuinely unrecoverable Documentation and metadata accuracy: narrowed the "zero engine changes" claim, corrected the theory-ordering protocol docstring, and set the development-status classifier to Beta CI dependency caps (numpy

Fields of Study & Subjects

Online Learning and Analytics Intelligent Tutoring Systems and Adaptive Learning Online and Blended Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 March 2026
Language
EN