Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Parallel Bayesian Optimization
Extension of Bayesian optimization where multiple points are evaluated simultaneously in parallel, reducing total optimization time by leveraging distributed computing resources.
Batch Acquisition
Strategy for selecting a batch of points to evaluate simultaneously, based on an acquisition function that optimizes a global performance criterion over the entire batch rather than a single point.
Batch GP-UCB
Variant of the GP-UCB (Gaussian Process Upper Confidence Bound) algorithm adapted for simultaneous selection of multiple points, by maximizing an upper bound on the sum of expected rewards from the batch.
Parallel Thompson Sampling
Batch acquisition method where multiple samples are drawn from the posterior distribution of the model, and the points corresponding to the maxima of these samples are evaluated in parallel.
Kriging Believer
Heuristic for sequential batch acquisition where each subsequent point is selected by assuming that previous evaluations in the batch are equal to their model predictions (the 'believer').
Local Penalization
Batch acquisition technique that penalizes the selection of points close to those already chosen in the current batch, by modifying the acquisition function to encourage spatial diversity of evaluations.
q-EI (q-Expected Improvement)
Generalization of Expected Improvement for batch optimization, calculating the expected improvement over a set of q points simultaneously, taking into account their joint correlation.
Max-Value Entropy Search (MVES)
Acquisition method that targets the reduction of uncertainty about the maximum value of the function, adapted for parallel evaluation by selecting points that maximize mutual information about this maximum.
Parallel Multi-Armed Bandit
Optimization framework where multiple arms (configurations) are pulled simultaneously at each round, applied to Bayesian optimization to accelerate exploration-exploitation with parallel evaluations.
Asynchronous BO
Bayesian optimization approach where the model is updated and new points are chosen as soon as an evaluation finishes, without waiting for all ongoing evaluations to complete, to maximize resource utilization.
Phantom Optimal
Concept in batch acquisition where a hypothetical optimal point is used to guide the selection of other points in the batch, penalizing points that would be optimal if this phantom existed.
Batch Knowledge Gradient (KG)
Extension of Knowledge Gradient for parallel optimization, evaluating the expected information gain of a batch of points on the final decision, taking into account correlations between these points.
Batch Acquisition Correlation
Measure of statistical dependence between acquisition function values for different candidate points in a batch, crucial for avoiding redundancy and maximizing the efficiency of parallel optimization.
Batch Diversity
Principle guiding the selection of batches of points to evaluate simultaneously, aiming to maximize spatial or informational dispersion of points to effectively cover the search space.
Internal vs External Parallelism
Distinction between internal parallelism (parallel evaluation of points within a single iteration) and external parallelism (execution of multiple sequential iterations in parallel on different resources).
Constant Liar Heuristic
Batch acquisition strategy where points already selected in the batch are assumed to have a constant value (e.g., the best observed value) when selecting subsequent points, to simplify computation.
GP-BUCB (Gaussian Process Batch Upper Confidence Bound)
Batch acquisition algorithm that extends the UCB principle to the simultaneous selection of multiple points, balancing exploration and exploitation at the batch level rather than at the individual point level.
Multi-Point Expected Improvement (MP-EI)
Variant of Expected Improvement that calculates the expected improvement for a set of points, often using approximations of the distribution of the maximum of these points for batch selection.
Parallel Sequential Kriging Optimizer (SKO)
Adaptation of the SKO optimizer for parallel evaluations, using strategies such as Kriging Believer or Local Penalization to build efficient batches of points.
Information-Theoretic Batch Acquisition
Class of batch acquisition functions based on information theory, such as Entropy Search or MES, which aim to maximize the information gained about the distribution of the global optimum from a batch of evaluations.