Show simple item record

dc.contributor.advisorHerdiyeni, Yeni
dc.contributor.advisorRidha, Ahmad
dc.contributor.authorAditya, Ahmad Yudha
dc.date.accessioned2026-09-17T08:28:43Z
dc.date.available2026-09-17T08:28:43Z
dc.date.issued2026
dc.identifier.urihttp://repository.ipb.ac.id/handle/123456789/179960
dc.description.abstractMateri Computational Thinking (CT) bersifat abstrak sehingga satu pendekatan pembelajaran seragam tidak efektif untuk semua mahasiswa. Personalized Learning System (PLS) yang ada masih statis, hanya membedakan dua tipe belajar dan tidak merespons dinamika pola belajar mahasiswa secara real-time. Data awal dari 451 mahasiswa peserta CT di IPB menunjukkan distribusi tipe kognitif yang tidak merata (90,0% tergolong PAR), mengindikasikan rekomendasi statis dua-tipe mengabaikan kelompok minoritas. Penelitian ini mengembangkan PLS adaptif berbasis Reinforcement Learning Contextual Bandit yang memperluas ruang aksi dari 2 menjadi 8 Learning Type dikombinasikan 6 level mastery, menghasilkan 48 kode kognitif. Optimasi hyperparameter dievaluasi pada lima metrik mutu pemodelan dan keberhasilan menemukan Learning Type, diuji terhadap 100.800 mahasiswa simulasi yang meliputi seluruh 48 kode kognitif. Konfigurasi Low Exploration (Epsilon_0=0,20, decay=0,995) terpilih sebagai terbaik dengan tingkat konvergensi 45,3% dan porsi pemilihan Learning Type sesuai sebesar 36,8%, dibandingkan 12,5% pada pemilihan acak maupun aturan statis. Keunggulan tersebut bertahan pada ketujuh skema variasi atribut, termasuk ketika kondisi awal mahasiswa tidak membawa informasi mengenai Learning Type yang sesuai. Verifikasi sesi nyata 20 pertanyaan mencapai akurasi 95%, dan survei terhadap 21 pengguna menunjukkan 85,7% merasakan kesesuaian dengan gaya belajar mereka. Integrasi RAG-FAISS dan LLM menghasilkan konten pembelajaran kontekstual selaras kode kognitif aktif.
dc.description.abstractComputational Thinking (CT) material is abstract, so uniform teaching is ineffective across students. Existing Personalized Learning Systems (PLS) remain static, distinguishing only two learning types and unresponsive to real-time changes in learning patterns. Preliminary data from 451 CT students at IPB show an uneven cognitive-type distribution (90.0% classified as PAR), indicating static two-type recommendations neglect minority learners. This research develops an adaptive PLS based on Contextual Bandit Reinforcement Learning, expanding the action space from 2 to 8 Learning Types combined with 6 mastery levels, producing 48 cognitive codes. Hyperparameter optimization was evaluated across five metrics of modelling quality and Learning Type identification, tested on 100.800 simulated students spanning all 48 cognitive codes. The Low Exploration configuration (Epsilon_0=0.20, decay=0.995) was selected as optimal, achieving a 45.3% convergence rate and selecting the appropriate Learning Type in 36.8% of questions, against 12.5% under both random selection and a static rule. The advantage persists across all seven attribute variation schemes, even under uninformative initial conditions. Verification on a real 20-question session achieved 95% accuracy, and a survey of 21 users showed 85.7% perceiving a match with their learning style. RAG-FAISS and LLM integration produces contextual content aligned with the student's active cognitive code.
dc.description.sponsorship
dc.language.isoid
dc.publisherIPB Universityid
dc.titlePENGEMBANGAN SISTEM REKOMENDASI TIPE BELAJAR MAHASISWA MENGGUNAKAN TEKNIK REINFORCEMENT LEARNING PADA PERSONALIZED LEARNING SYSTEMid
dc.title.alternative
dc.typeSkripsi
dc.subject.keywordContextual Banditid
dc.subject.keywordEpsilon-Greedyid
dc.subject.keywordPersonalized Learning Systemid
dc.subject.keywordTipe Belajarid
dc.subject.keywordReinforcement Learningid
dc.subtypeUndergraduate Theses


Files in this item

Thumbnail
Thumbnail
Thumbnail

This item appears in the following Collection(s)

Show simple item record