Multi-objective reward generalization: improving performance of Deep Reinforcement Learning for applications in...

Area of Science:

• Computational Finance
• Artificial Intelligence
• Machine Learning

Background:

• Traditional single-asset trading strategies often struggle with dynamic market conditions and sparse reward signals.
• Deep Reinforcement Learning (DRL) offers potential for automated trading but requires careful reward function and discount factor specification.
• Multi-objective optimization can provide a more robust framework for complex decision-making processes in finance.

Purpose of the Study:

• To investigate the efficacy of a Multi-Objective Deep Reinforcement Learning (MODRL) algorithm for single-asset trading in both stock and cryptocurrency markets.
• To evaluate the algorithm's ability to generalize reward functions and discount factors within the learning process.
• To compare the performance and stability of the MODRL approach against traditional Single-Objective (SO) DRL strategies.

Main Methods:

• Development and implementation of a novel MODRL algorithm that incorporates reward generalization and discount factor learning.
• Empirical testing on diverse financial assets including cryptocurrencies (BTCUSD, ETHUSDT, XRPUSDT) and stocks (AAPL, SPY, NIFTY50).
• Comparative analysis of MODRL against SO strategies, focusing on predictive stability and performance under sparse reward conditions.

Main Results:

• The MODRL algorithm demonstrated superior reward generalization capabilities compared to SO approaches.
• Preliminary statistical evidence indicates enhanced predictive stability for the MODRL strategy across various assets.
• The MODRL algorithm exhibited a significant performance advantage in scenarios characterized by sparse reward mechanisms.

Conclusions:

• The proposed MODRL framework offers a promising advancement for automated single-asset trading by inherently handling complex reward structures.
• The algorithm's ability to generalize reward functions and discount factors contributes to increased predictive stability and robustness.
• Open-source code availability facilitates further research and application of MODRL in financial trading.