Authors:
Ryoga Miyajima
and
Katsuhide Fujita
Affiliation:
Tokyo University of Agriculture and Technology, Koganei, Tokyo, Japan
Keyword(s):
Automated Negotiation, Concurrent Negotiation, Multi-Agent System, Supply Chain Management, Reinforcement Learning, Representation Learning.
Abstract:
In the field of automated negotiation, significant attention has been paid to methods for learning negotiation strategies using reinforcement learning. However, in concurrent negotiation, where negotiation proceeds with multiple counterparties with various strategies in parallel, it is difficult to consider the differences in the strategies of counterparties using the conventional formulation in which the state is defined using the bids of both counterparties. In this study, we propose a reinforcement learning framework for learning negotiation strategies that considers the strategy models of the negotiation partners in concurrent negotiations. Strategy modeling is realized using embeddings with a representation function based on the unsupervised learning of generative–discriminative representations from negotiation log data. Through evaluation experiments, we show the performance of the representation function in identifying the strategies of negotiation partners and the effectivene
ss of introducing the representation function into the reinforcement learning of negotiation strategies.
(More)