Heterogeneous Network Harvesting for Budgeted Workload Placement

The patent applies partial action by performing selective harvesting that queries only a subset of nodes in the heterogeneous network rather than conducting exhaustive searches. The system iteratively queries nodes based on predicted utility scores, stopping when the budget is exhausted or when confidence in the selected node is sufficient, thus achieving acceptable precision without the time cost of complete exploration.

The system performs preliminary actions by pre-computing node embeddings and training the prediction model before the actual workload placement query. This allows the system to quickly evaluate candidate nodes during the harvesting phase without performing expensive real-time computations, thereby reducing the time required for node selection while maintaining accuracy.

The patent implements partial action by querying only the necessary subset of nodes to achieve satisfactory placement accuracy. The selective harvesting process stops when the budget constraint is reached or when the predicted utility score indicates sufficient confidence, avoiding the energy waste of querying all nodes in the network while still identifying suitable candidates.

The system uses feedback mechanisms by iteratively querying nodes, updating the heterogeneous network model with observed data, and re-evaluating candidate nodes based on new information. This allows the system to adapt its querying strategy dynamically, focusing energy on promising candidates rather than uniformly querying all nodes, thus improving energy efficiency while maintaining accuracy.

The patent introduces node embeddings as an intermediary representation that captures complex heterogeneous network structures and constraints in a compressed vector form. These embeddings serve as mediators between the complex HN structure and the prediction model, allowing the system to leverage detailed constraint modeling without directly processing the full complexity of the heterogeneous network during query time.

The system applies parameter changes by transforming the heterogeneous network into a different parameter space using node embeddings. This transformation converts complex structural relationships into continuous vector representations that are easier to process and compare, reducing the effective complexity while preserving the essential constraint information needed for accurate node selection.

The patent applies partial action by performing iterative network expansion only to the extent necessary to discover useful information within the budget constraint. The selective harvesting process expands the heterogeneous network incrementally, querying nodes in order of predicted utility, and stops when the budget is exhausted or when sufficient information has been gathered to make an informed decision, avoiding unnecessary exploration time.

The system performs preliminary actions by pre-training the prediction model on available data before the iterative expansion phase. This allows the model to make informed predictions about which nodes to query next, guiding the iterative expansion process more efficiently and reducing the time required to discover useful information compared to random or systematic exploration methods.