Evaluating Stream-Learning Models Under Strict Embedded Memory Constraints

Across 6,463 experimental runs, the authors document how memory consumption diverges between adaptive ensembles and incremental tree-based methods during continuous operation. The results carry direct implications for deploying online learners on microcontroller-class hardware where RAM ceilings are fixed.

Benchmark Configuration

Seven stream classifiers were evaluated against 13 streams, combining real-world and synthetic sources. Each classifier was constrained to a target model-size budget spanning four orders of magnitude, from 128 KiB at the low end to roughly 8 MiB at the upper limit. The total of 6,463 experiments provides dense coverage of the trade-off surface between predictive performance and memory footprint under resource constraints.

Observed Memory Dynamics

Adaptive ensembles exhibited budget violations almost immediately, exceeding the smaller memory caps during early-stream adaptation as base learners were added. Incremental trees, by contrast, displayed delayed growth: within budget at deployment, but accumulating nodes over time and expanding substantially during long-running streams. The divergence establishes a temporal axis to memory budgeting that static footprint analyses miss entirely.

Practical Implications

For embedded deployment, the benchmark signals that ensemble-based online learners require overprovisioned RAM or aggressive pruning to remain within tight constraints. Incremental tree approaches offer compact initial states but introduce non-stationary memory demands that complicate worst-case certification. Practitioners targeting constrained devices should treat model-size not as a deployment-time constant but as a variable governed by stream duration and drift exposure.