Hybrid Intelligent Systems for Adaptive Problem Solving in Dynamic Environments

Abstract

Hybrid intelligent systems combining symbolic reasoning with subsymbolic learning have emerged as a compelling approach to adaptive problem
solving in environments characterised by incomplete information, non-stationarity, and conflicting objectives. While purely neural architectures excel
at pattern recognition, they lack the structural compositionality required for robust generalisation across dynamically shifting problem spaces. This
study presents AdaptHIS, a hybrid intelligent framework that tightly couples a neural feature extractor with a symbolic constraint-satisfaction engine
and a reinforcement learning policy network to enable context-aware, adaptive decision-making. AdaptHIS was evaluated across four canonical
dynamic environment benchmarks: OpenAI Gym MountainCar-Continuous, LunarLander-v2, a real-time resource scheduling simulation (n = 180,000
episodes), and a multi-agent traffic signal control environment modelling The Hague urban network (62 intersections). AdaptHIS achieved cumulative
reward improvements of 27.4%, 31.8%, 22.6%, and 34.1% respectively over state-of-the-art baselines including PPO, SAC, and DQN. Symbolic
constraint integration reduced policy violations by 64.3% compared to purely neural agents. Convergence speed improved by 41.7% on average
across all benchmarks. These results confirm that neurosymbolic hybridisation substantially enhances adaptability, safety, and efficiency in dynamic
intelligent systems.