Abstract: Aiming at the automated negotiation problem, a co-training based semi-supervised opponent's negotiation preference learning method was proposed. In this method, negotiation process was mapped into two new feature spaces：price orbit feature space and interaction orbit feature space. Two support vector regression machines were trained in their feature space respectively, and confident labeled instances for each other alternately were provided, thus the scale of training samples was extended. The opponent's negotiation preference was obtained by two machines. Win-win negotiation counter proposal based on both sides'negotiation preference was proposed by negotiation decision model. Experiment results showed that the proposed method could improve total negotiation utility, reduce negotiation round and save negotiation time.

Key words: negotiation model, semi-supervised learning method, support vector regression machine, preference, simulated annealing algorithm

摘要： 针对自动化协商问题，提出一种基于协同训练的半监督对手协商偏好学习方法。在该方法中，将协商过程映射到出价轨迹特征空间和交互轨迹特征空间两个新的特征空间。在两个特征空间中分别训练支持向量回归机，两个学习机迭代，互相提供可靠的有标记训练样本，以扩大训练样本规模。由两个学习机共同学习，得到对手的协商偏好。协商决策模型以双方协商偏好为基础提出双赢的协商反建议。实验数据表明，所提方法可以提高协商总体效用，减少协商回合数，节省协商时间。

关键词: 协商模型, 半监督学习方法, 支持向量回归机, 偏好, 模拟退火算法