A Policy Gradient Based Particle Swarm Optimizer for Portfolio Optimization Problem

Due to complexity and uncertainty of financial market, how to reasonably distribute wealth amongst different assets to maximize return or minimize risk has long been a challenging research topic, which is usually called the portfolio optimization problem (POP). To deal with such difficulties, this paper proposes a policy gradient based particle swarm optimizer (PG-PSO). By combining PSO with policy gradient algorithm in reinforcement learning, the proposed method provides a novel PSO parameters adjustment mechanism which improved the optimization accuracy while at the same time reducing the workloads of manually configuring parameters. A policy neural network as an agent is constructed and interacts with the particle swarm to realize the adaptive update of PSO parameters. According to the test results of several typical benchmark functions, it can be seen that PG-PSO has better performance compared with other algorithms mentioned in this paper. The experimental results of POP show that PG-PSO can also improve Sharpe ratio value of the entire portfolio more effectively.