Study finds evolution strategies are simpler but underperform deep reinforcement learning, offering limited benefits as a pretraining method

Evolution strategies (ES) provide a derivative-free, computationally cheaper alternative to Deep RL but generally fail to match its performance levels

Researchers tested ES and DRL across varying difficulty levels including Flappy Bird, Breakout, and Mujoco environments to compare training efficiency

ES showed no consistent speed advantage over DRL and only provided benefits in simpler environments like Flappy Bird when used for pretraining

Results suggest ES are unsuitable for complex decision-making tasks, raising questions about their viability for demanding real-world scenarios