Skip to content
Open access

A Self-Adaptive Deep Learning Architecture for Autonomous Intelligence and Decision-Making in Complex Computational Environments

Sep 2026 · Computer and Decision Making: An International Journal · 0 citations

Abstract

Intelligent systems that can adapt their decisions to changes in operating conditions are needed in dynamic computational environments, and conventional predictive models could fail if the states in the environment and the utility of the decisions change. This study developed and evaluated a reward-adaptive deep learning architecture for autonomous action selection under heterogeneous computational conditions. Sequential state, action, reward and throughput records were used and a quantitative computational design was applied. The benchmark performance was analysed in the context of DARA, ID and MIN policies, and the state-dependent decision behaviour was analysed with correlation and non-parametric statistical tests. A baseline deep neural network was compared with an architecture where the reward is adapted based on the training/validation/testing data, obtained from separate scenarios. Reward and autonomous action selection were highly related to the environmental state. The reward-adaptive model improved test accuracy from 0.5427 to 0.5946, macro F1 from 0.2876 to 0.3543, and weighted F1 from 0.5571 to 0.5763, although balanced accuracy decreased from 0.4414 to 0.3770. The mean expected reward was 8.80% higher than the baseline model. Two unseen scenarios showed improvement in generalization performance and one showed a decrease, indicating scenario-dependent adaptation. The results show that reward-guided deep learning can improve autonomous decision-making, both predictive and utility, and that it is still essential to have strong evaluation over heterogeneous states and action distributions.  

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.