Prediction. Past work has shown that cloud workloads can be highly variable and may not be easily described by single well- known distributions [11]. To address this problem we run a Xxxxx-Xxxxx-style simulation on-line to generate the empirical distribution of expected spot instance lifetimes. However, we note that the time-until-eviction is affected by the capacity of the cloud that is occupied by un-evictable on-demand workload and other spot instances. Intuitively, if the cloud is relatively “empty”, a spot-instance that is introduced will likely live longer than if the cloud is close to “full” capacity. Thus, our Xxxxx- Xxxxx simulation produces a set of empirical distributions, one conditioned on each level of possible occupancy. For example, a cloud with 100 cores has 101 possible occu- pancy levels: from 0 cores occupied to 100 cores occupied and each level of occupancy corresponds to a different distribution of spot-instance lifetimes. We use quantiles of these distributions to quote the expected lifetime to the scheduler during the admission control phase based on the current occupancy level at the time the spot-instance request is made. If the instance (based on its maximum lifetime specified by its user) is expected to be evicted with a higher probability than specified by the target probability (quoted as an SLA) for the cloud, it is rejected (not admitted). The cloud administrator is responsible for setting the SLA on eviction probability that is advertised to all cloud users. The Xxxxx-Xxxxx simulator generates a sample of “fictitious” spot-instance requests that using the recent cloud load history. It repeatedly chooses a random point in the history and simu- lates the arrival and eviction of a spot-instance, recording the occupancy level at he time the spot-instance starts and its time- until-eviction. Running faster than real time, it generates 10000 such samples and divides them into empirical distributions based on occupancy level.
Prediction. Several supervised learning approaches were investigated, and RF performed the most accurately and robustly among all the approaches. Hence, results were demonstrated using RF, which was achieved with R package randomForest (Liaw & Xxxxxx, 2002). In the RF, a binary classification model was trained using methylation score together with genomic features. In each trained model, the importance of input features was assessed using Gini gain importance. The number of trees used in the model was determined by the stability of out-of-bag error. The predicting result is represented by the probability of getting a vote from the randomly generated classification tree for each class. The predicting performance was evaluated using the ChIP-seq peaks as the gold standard. ROC based on the class probability and the gold standard was computed to show the overall predicting performance of our method.
Prediction. For each cell-TF combination to be predicted in the ladder stage, prediction was made using all the trained models from all training cell types. For one TF with 𝑚 training cell types, prediction results from 10 models within each training cell type is averaged; then the weighted sum of these 𝑚 averaged result were used as the final prediction score. The weight is assigned based on the local similarity of DNase profile. This is based on the motivation that similar chromatin landscape might hint similar functionality during regulation. DNase profile similarity is calculated as the correlation coefficient between the tested cell type and the 𝑚 training cell types. For example, In the case of ATF2 binding prediction within K562. Firstly, input matrix is constructed by collecting: (1) ATF2 motif sets matrix, (2) DNase matrix from K562 DNase-seq data, (3) RNA matrix from K562 RNA-seq data. Then the input matrix is fed into all the 30 trained models, resulting in 30 prediction scores for every predicting region. Prediction scores within the same training cell are averaged, resulting in 3 scores: GM12878 score, H1-hESC score and MCF-7 score. Finally, these three scores are weighted average based on the DNase profile similarity between the training cells and K562. To be more specific: the more similar the local DNase landscape within this region between GM12878 and K562, the more weight the GM12878 score will get. Prediction for the final round adopts the same strategy, except that applying on a larger test sets.
Prediction. The task of time-series prediction has to do with forecasting (typically) future values of the time series based on its past samples [96]. For this purpose, we need to build a predictive model for the data. The autoregressive family of models can be used to predict a future value as a linear combination of earlier sample values, provided the time series is stationary. Linear non- stationary models like ARMA models have also been found useful in many economic and industrial applications where some suitable variant of the process can be assumed to be stationary. Another popular work-around for non-stationarity is to assume that the time series is piece-wise stationary. The series is then broken down into smaller pieces called “frames”, within each of which the stationary condition can be assumed to hold, and then separate models are learnt for each frame. In addition to this standard ARMA family of models, there are many nonlinear models for time series prediction, e. g. neural networks. The prediction problem for symbolic sequences has been addressed in Artificial Intelligence research, regarding various rule models, such as disjunctive normal form model, periodic rule model etc. Based on these models sequence-generating rules are obtained that state some properties that constrain which symbol can appear next in the sequence. In many cases, prediction may be formulated as classification, association rule finding or clustering problems. Generative models can also be used effectively to predict the evolution of time series.
Prediction. Standard errors and confidence intervals for the survivor function from the estimator can be obtained using the Delta method (Casella and Xxxxxx 2002). On investigation the method was complex and for many repeated measures would not support the prediction. Therefore, I recommend a bootstrap routine to estimate and present uncertainty surrounding the function (Efron and Tibshirani 1993). Illustration of this shown in Section 4.6.1.
Prediction. Prediction implies that the model provides insights in how the real world system is behaving. Ultimately this implies that the value of a prediction can only be esti- mated if the model results can be compared to empirical data. Assuming that such data are available, we confront a fundamental problem, as we have to realize that – given a complex behaving system – empirical data reflect only one outcome (or market history) of a possible and unknown wide distribution of outcomes. Espe- cially if the empirical data reflect a situation which is highly unlikely to happen – which would be difficult to assess – the results obtained with a correct agent- based model would not match the available data1. Obviously, a valid model should be capable of replicating the empirically observed phenomena, and hence over a large number of simulation runs a limited number of runs should replicate the empirical data. However, here we still have the situation that a majority of runs provides outcomes that differ from the empirical data. Because the real system might have looked rather different, point prediction of agent-based models remain at least a risky business. A solution that is frequently being mentioned is the gen- eration of outcome distributions using a simulation model. This is a valuable method in exploring if the empirical data fit within the range of model outcomes. However, considering the qualitatively very different outcomes that may emerge in social complex systems, such distributions cannot be interpreted as normal dis- tributions, thus excluding the possibility of determining if the empirical data fit within a certain confidence interval of outcomes. However, rather than focusing at precise predictions, e.g. of the distribution of certain activities, it would also be possible to describe the system behaviour in more abstract terms. Under such assumptions, we propose an approach using a generic measurement of the volatility of the system, which can be expressed in terms of the variance in system behaviour over time. The aim of this approach is to capture more abstract attributes of the system, such as the dynamics of change. If a simulation tool is capable of generating system dynamics that closely reflect the dynamics as ob- served in empirical situations (e.g., the dynamics of change – social transitions), it can be stated that the simulation model is capable of predicting system properties, and can be used to explore how certain policies might interfere wi...
