Hierarchical reinforcement learning for trading agents

Abstract

Autonomous software agents, the use of which has increased due to the recent growth in computer power, have considerably improved electronic commerce processes by facilitating automated trading actions between the market participants (sellers, brokers and buyers). The rapidly changing market environments pose challenges to the performance of such agents, which are generally developed for specific market settings. To this end, this thesis is concerned with designing agents that can gradually adapt to variable, dynamic and uncertain markets and that are able to reuse the acquired trading skills in new markets. This thesis proposes the use of reinforcement learning techniques to develop adaptive trading agents and puts forward a novel software architecture based on the semi-Markov decision process and on an innovative knowledge transfer framework. To evaluate my approach, the developed trading agents are tested in internationally well-known market simulations and their behaviours when buying or/and selling in the retail and wholesale markets are analysed. The proposed approach has been shown to improve the adaptation of the trading agent in a specific market as well as to enable the portability of the its knowledge in new markets.