Abstract
<jats:p>The article investigates the impact of fully connected neural network configuration on the performance of the Deep Q-learning algorithm in the CartPole-v1 environment. Special attention is given to the trade-off between learning quality and computational efficiency. Three configurations of a fully connected neural network were considered: [64], [128, 128], and [256, 256, 128]. Each configuration was evaluated over 20 independent runs with a training duration of 500 episodes. The evaluation was based on the following metrics: training time, execution speed (steps per second), memory usage, average reward, and number of episodes to convergence. The results show that increasing model complexity leads to higher memory consumption and lower execution speed. The simplest configuration demonstrated the highest computational efficiency but lower learning performance, while the most complex configuration achieved the highest reward at the cost of significantly increased resource usage. The configuration [128, 128] was found to provide the best balance between learning quality and computational efficiency. Keywords: reinforcement learning; Deep Q-learning; fully connected neural network; model configuration; performance; computational efficiency; CartPole-v1.</jats:p>