.png){kind=icon}
.png){kind=icon}
Return
Battery-Adaptive Collaborative Inference via Hybrid Deep Reinforcement Learning in Heterogeneous Edge Intelligence
.png){kind=icon}
PRE
AI
H
H
X
王
M
DOI：10.1109/jiot.2026.3704290.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid growth of edge intelligence demands efficient and sustainable deep neural network inference on battery-constrained mobile devices. Many existing studies treat the battery state as static or short-term and overlook its long-horizon evolution. This leads to limited adaptability and degraded device survivability, especially in battery-critical regimes. We study battery-aware collaborative inference in heterogeneous environments and formulate a multiobjective joint optimization that minimizes end-to-end latency and energy consumption while maximizing inference accuracy. The resulting problem is non-deterministic polynomial (NP) -hard and can be formulated as a mixed-integer nonlinear program. To tackle it, we propose a hybrid two-stage attention soft actor–critic (HTASAC), a hybrid two-stage attention-based soft actor–critic framework. In the first stage, we develop a structure-aware discrete policy network that integrates cross-attention and low-rank bilinear interactions to explicitly model the coupling between model partitioning and early-exit selection. This design captures high-order dependencies and improves decision efficiency in large discrete action spaces. In the second stage, we embed a karush-kuhn-tucker (KKT)-based convex optimizer to derive closed-form solutions for dynamic frequency scaling, enabling physically consistent resource scheduling without quantization-induced errors. Extensive simulations demonstrate the effectiveness of HTASAC over representative state-of-the-art baselines. In battery-critical scenarios, HTASAC reduces inference latency by 27.6% and further decreases system energy consumption by 11.3% compared with the runner-up hybrid soft actor-critic (HSAC), thereby prolonging the operational lifespan of local devices.
Keywords:
Battery-aware optimization
collaborative inference
convex optimization
edge intelligence
hybrid deep reinforcement learning
Journal
IF:
8.9
Papers:
1.4W
Citations:
7.8W
Organization
Cited Papers