Enterprise integration platforms often synchronize external-system data at fixed intervals. Such schedules do not adapt to activity: they produce empty synchronizations during quiet periods and stale data during bursts. This thesis studies adaptive synchronization scheduling with reinforcement learning (RL), with selected configurations using forecasts of events per minute (ER) or time to the next active minute (TTE). We evaluate synchronization policies on eight units across four platforms. For Brightpath Salesforce, we evaluate generalization across a regime break in activity. On frozen test traces with synchronized initial states, we construct Pareto frontiers for synchronization burden and tail staleness and compare RL agents with a predefined set of periodic intervals at one evaluated phase per interval. RL contributes at least one non-dominated operating point for every unit, but under the primary logarithmic specification periodic policies cover more sampled preferences on average; some selections change under linear scaling. Descriptive comparisons and a limited matched ablation did not show a consistent general benefit from forecast features. The available comparisons also do not establish a reliable relationship between forecast accuracy and scheduling value. Policy selection must therefore be validated for each unit and its operational priorities.
|
