以量子遺傳演算法與深度學習為基礎之自動化協商對手喜好預測

電子化協商(e-Negotation)可透過網路讓協商者免除空間及時間的限制，大幅節省協商成本。此外，透過協商支援系統，可協助協商者以更理性的態度快速做出決策，大幅降低過程中的誤判。然而，面對日益複雜的交易環境，協商者仍需親自參與許多流程，使其便利性大打折扣。有鑑於此，學者便開始發展各種有效的自動化協商機制。前人研究雖已提出許多可行的做法，但在機器學習領域日新月異的年代，實有必要重新評估現有架構，發展更準確、更符合實際需求之協商預測方法。為此，本研究以Faratin等人(1998)提出之協商模型為基礎，透過量子遺傳演算法與深度學習方法，發展有效的協商對手喜好預測方法。首先，我們針對協商中各種參數進行基因編碼，以表示雙方的喜好。接下來，運用量子遺傳演算法找尋對手喜好可能對應之基因組合，並透過量子旋轉閘角度的調整，產生更合適的預測結果。其次，我們也以雙方之出價序列為輸入，設計與時序相關之LSTM無模型對手喜好預測方法，並與前述方法進行比較。為驗證提出方法之有效性，本研究透過模擬協商進行效能評估。實驗結果顯示，與傳統的基因演算法與粒子群演算法比較，無論在整合式或分配式協商中，本研究使用之預測方法均可獲得較佳之總效用。然而，在使用LSTM進行無模型的預測時，並無法獲得合理的預測結果。綜合上述結果，我們發現以喜好模型為基礎之量子遺傳預測方法，確實有助於進一步提升雙方總效用，產生雙贏的協商結果。

關鍵字

協商對手喜好預測；量子遺傳演算法；基因演算法；粒子群演算法； LSTM ；電子化協商；電子商務

並列摘要

Electronic negotiation (e-negotiation) allows traders to conduct negotiations through the Internet without considering the time-zone and physical location, that will greatly reduce the negotiation costs. In addition, the negotiation support system can help negotiators make quick decisions with a more rational attitude that would reduce misjudgments in negotiation. However, to face an increasingly complex trading environment, negotiators still need full participation that could degrade the convenience of e-negotiation seriously. In the light of this, scholars have begun to develop effective automated negotiation mechanisms. Although previous studies have proposed many feasible methods, with the quick development of machine learning, it is still necessary to re-evaluate the existing architecture and develop more suitable negotiate support methods. For this reason, this research is based on the negotiation model proposed by Faratin et al. (1998), developing an effective method for predicting the preferences of negotiating opponent through Quantum-Inspired Genetic Algorithms and Deep Learning. First, each gene is encoded by various parameters in the negotiation to express the preferences of both parties. Secondly, the quantum-inspired genetic algorithms is used to find the genetic combination corresponding to the opponent's preferences, through the adjustment of the angle of the quantum rotation gate. Also, to compare with the aforementioned methods, we take the bid sequences of both parties as input and design a time-related of LSTM for model-free opponent preference prediction. To verify the effectiveness of the proposed method, this study conducts simulated negotiation to compare the obtained results. Comparing with traditional genetic algorithm and particle swarm optimization, the proposed method can achieve better overall utility no matter in integrated or distributed negotiation. However, when LSTM is used for model-free prediction, reasonable prediction results cannot be obtained. In summary, we found that the preference prediction methods based on quantum-inspired genetic algorithm does help to further enhance the total utility of both parties and reach a win-win negotiation result.

並列關鍵字

preference prediction in negotiation ； quantum-inspired genetic algorithms ； genetic algorithms ； particle swarm optimization ； lstm ； electronic negotiation ； e-commerce