Glosarium AI
Kamus lengkap Kecerdasan Buatan
Real-Time Acquisition Function
Strategy for selecting evaluation points in Bayesian optimization, designed to minimize decision latency by dynamically adapting the acquisition criterion to time constraints.
Incremental Surrogate Model
Substitute model (often a Gaussian process) updated incrementally with each new observation, avoiding complete retraining to ensure real-time performance.
Real-Time Budget
Strict time constraint imposed on each iteration of Bayesian optimization, limiting the time allocated to model updating and acquisition criterion calculation.
Low-Computational-Cost Acquisition Criterion
Acquisition function (such as expected improvement or upper confidence bound) simplified or approximated to drastically reduce computation time, essential for fast optimization loops.
Online Gaussian Process
Gaussian process implementation where inference and prediction are performed continuously as data arrives, without requiring batch processing.
Dynamic Search Space Pruning
Technique to reduce the candidate space for evaluation at each iteration based on accumulated information, to accelerate the selection of the next point.
Streaming Bayesian Optimization (Streaming BO)
Approach where optimization is performed on a continuous data stream, adapting the model and optimization decisions in real-time as new information arrives.
Meta-Learning for BO Acceleration
Use of meta-learning to quickly initialize the surrogate model with knowledge from past optimization tasks, thereby reducing the number of iterations needed in real-time.
Data Parallelism for Real-Time Bayesian Optimization
Parallelization strategy where computations related to the surrogate model (e.g., predictions on candidate point sets) are distributed to meet strict timing constraints.
Sliding Window of Observations
Method where only a recent subset of observations is retained for model updates, limiting computational complexity to ensure consistent real-time performance.
Episodic Bayesian Optimization
Framework where optimization is broken down into short temporal episodes, each with its own surrogate model, allowing frequent resets to adapt to dynamic environments.
Random Least Squares Approximation for Bayesian Optimization
Kernel approximation technique in Gaussian processes using random features to reduce matrix inversion complexity from O(n^3) to O(n*m), where m << n.
Constant Time Prediction
Objective of real-time Bayesian Optimization implementations where the latency for predicting the surrogate model's mean and variance for a new point is guaranteed to be below a fixed threshold.
Closed-Loop Control System with Bayesian Optimization
Application of real-time Bayesian Optimization as a component of a control system, where actions are continuously adjusted based on model predictions to optimize performance.
Active Real-Time Optimization
Process where the Bayesian Optimization algorithm decides not only the next point to evaluate but also the optimal timing for this evaluation to maximize information gain under time constraints.
Sparse Surrogate Model
Use of surrogate models (such as sparse Gaussian processes or Gaussian processes with inducing points) that exploit sparse data structures for fast updates and predictions.
Non-Myopic Acquisition Criterion
Criterion that evaluates the potential of a point over multiple steps ahead, often approximated to be computable in real-time, in order to make more robust decisions than myopic criteria.