Evolutionary DRL Environment: Transfer Learning-Based Genetic Algorithm

Abstract

Stock markets trading has risen as a critical challenge for artificial intelligence research. The way stock markets are moving and changing pushes researchers to find more sophisticated algorithms and strategies to anticipate the market movement and changes. From the artificial intelligence perspective, such environments require artificial agents to coordinate and transfer their best experience through different generations of agents. However, the existing agents are trained using hand-crafted expert features and expert capabilities. Notwithstanding these refinements, no previous single system has come near to dominating the trading environment. We address the algorithmic trading problem utilising an evolutive learning method. Precisely, we train a multi-agent reinforcement learning algorithm that uses only self trades generated by different generations of agents. The evolution-based genetic algorithm operates as an evolutive environment that continually adapts the agent's internal strategies and tactics. Also, it pushes the system forward to generate creative behaviours for the next generations. Additionally, a deep recurrent neural network drives the mutation mechanism through the attention that dynamically encodes the memory mutation size. The winner, which is the last agent, achieved promising performances and surpassed traditional and intelligent baselines.