// Thompson Sampling - Multi-Armed Bandit (C++)
// Netflix 썸네일 최적화와 동일한 알고리즘

class ThompsonSamplingBandit {
    struct Arm {
        std::string name;       // e.g., "thumbnail_A"
        double alpha = 1.0;    // Beta 분포 성공 파라미터
        double beta = 1.0;     // Beta 분포 실패 파라미터
        int64_t total_pulls = 0;
        double total_reward = 0.0;
    };

    std::vector<Arm> arms_;
    std::mt19937 rng_;

public:
    // 최적의 arm 선택 (Thompson Sampling)
    int select_arm() {
        int best_arm = 0;
        double best_sample = -1.0;

        for (int i = 0; i < arms_.size(); i++) {
            // Beta 분포에서 샘플링
            std::gamma_distribution<double> gamma_a(arms_[i].alpha, 1.0);
            std::gamma_distribution<double> gamma_b(arms_[i].beta, 1.0);
            double a = gamma_a(rng_);
            double b = gamma_b(rng_);
            double sample = a / (a + b);

            if (sample > best_sample) {
                best_sample = sample;
                best_arm = i;
            }
        }
        return best_arm;
    }

    // 결과 피드백 (보상 업데이트)
    void update(int arm_id, double reward) {
        arms_[arm_id].alpha += reward;
        arms_[arm_id].beta += (1.0 - reward);
        arms_[arm_id].total_pulls++;
        arms_[arm_id].total_reward += reward;
    }

    // 현재 최적 arm의 추정 승률
    double best_estimated_rate() {
        double best = 0.0;
        for (auto& arm : arms_) {
            best = std::max(best, arm.alpha / (arm.alpha + arm.beta));
        }
        return best;
    }
};