Prediksi Multi-Langkah Perilaku Lawan dalam Interaksi Closed-Loop pada Iterated Prisoner's Dilemma

Faqih Mahardika, Dr. Sri Mulyana, M.Kom.

2026 | Skripsi | ILMU KOMPUTER

Prediksi perilaku lawan merupakan komponen penting dalam opponent modeling pada lingkungan interaksi berulang. Prediksi satu langkah yang digunakan secara rekursif memungkinkan peramalan multi-langkah, tetapi berpotensi menurunkan kualitas prediksi akibat akumulasi kesalahan. Penelitian ini bertujuan mengembangkan model prediksi trajektori aksi lawan pada permainan Iterated Prisoner's Dilemma serta mengevaluasi performanya pada horizon multi-langkah.

Penelitian menggunakan dataset hasil simulasi turnamen yang melibatkan 68 strategi dari pustaka Axelrod dengan total 23.120 permainan. Model dikembangkan menggunakan arsitektur Long Short-Term Memory (LSTM) dan dievaluasi melalui strategi pelatihan pembobotan loss sekuensial menggunakan Rank Order Centroid (ROC) serta dua mekanisme prediksi, yaitu Rollout Berbasis Model dan Rollout Berbasis Strategi.

Hasil penelitian menunjukkan bahwa model mencapai F1-score rata-rata sebesar 0,86 pada data pengujian. Rollout Berbasis Model menghasilkan F1-score rata-rata sebesar 0,885, namun mengalami penurunan sebesar 0,54 poin persentase pada setiap peningkatan satu horizon prediksi. Sebaliknya, Rollout Berbasis Strategi memperoleh F1-score rata-rata sebesar 0,879 dengan laju penurunan yang lebih kecil, yaitu 0,20 poin persentase per horizon, sehingga menunjukkan stabilitas prediksi yang lebih baik. Hasil ini menunjukkan bahwa pendekatan LSTM efektif untuk peramalan trajektori aksi multi-langkah, dengan Rollout Berbasis Model memberikan akurasi lebih tinggi, sedangkan Rollout Berbasis Strategi menawarkan stabilitas prediksi yang lebih baik.

Opponent behavior prediction is an important component of opponent modeling in repeated interaction environments. Recursively applying one-step predictions enables multi-step forecasting; however, this approach may lead to error accumulation, resulting in decreased prediction quality over longer forecasting horizons. This study aims to develop an opponent action trajectory prediction model for the Iterated Prisoner's Dilemma and evaluate its ability to maintain prediction accuracy across multi-step forecasting horizons.

The study utilizes a tournament simulation dataset involving 68 strategies from the Axelrod library, comprising a total of 23,120 games. The prediction model is developed using a Long Short-Term Memory (LSTM) architecture and evaluated through sequential loss-weighting training based on the Rank Order Centroid (ROC), as well as two forecasting mechanisms, namely Model-Based Rollout and Strategy-Based Rollout, to predict opponent actions up to ten steps ahead at each game turn.

The results show that the proposed model achieves an average F1-score of 0.86 on the test dataset. Performance analysis across forecasting horizons indicates that the Model-Based Rollout achieves a higher average F1-score of 0.885 but experiences an average degradation of 0.54 percentage points for each additional forecasting horizon. In contrast, the Strategy-Based Rollout achieves an average F1-score of 0.879 with a smaller degradation rate of 0.20 percentage points per horizon, demonstrating greater prediction stability over longer forecasting horizons. These findings indicate that the proposed LSTM-based approach is effective for multi-step opponent action trajectory forecasting in repeated interaction environments, with the Model-Based Rollout providing higher predictive accuracy while the Strategy-Based Rollout offers better stability as the forecasting horizon increases.

Kata Kunci : pemodelan lawan, permainan berulang, prediksi multi-langkah, interaksi closed-loop, Iterated Prisoner's Dilemma