H2O Driverless AI is an artificial intelligence (AI) platform for automatic machine learning. Driverless AI automates some of the most difficult data science and machine learning workflows such as feature engineering, model validation, model tuning, model selection and model deployment. It aims to achieve highest predictive accuracy, comparable to expert data scientists, but in much shorter time thanks to end-to-end automation. Driverless AI also offers automatic visualizations and machine learning interpretability (MLI). Especially in regulated industries, model transparency and explanation are just as important as predictive performance. Modeling pipelines (feature engineering and models) are exported (in full fidelity, without approximations) both as Python modules and as pure Java standalone scoring artifacts.

This section provides answers to frequently asked questions. If you have additional questions about using Driverless AI, post them on Stack Overflow using the driverless-ai tag at http://stackoverflow.com/questions/tagged/driverless-ai.


How is Driverless AI different than any other black box ML algorithm?

Driverless AI uses many techniques (some older and some cutting-edge) for interpreting black box models including creating reason codes for every prediction the system makes. We have also created numerous open source code examples and free publications that explain these techniques. Please see the list below for links to these resources and for references for the interpretability techniques.

  • Open source interpretability examples:
  • Free Machine Learning Interpretability publications:
  • Machine Learning Techniques already in Driverless AI:


How can I change my username and password?

The username and password is tied to the experiments you have created. For example, if I log in with the username/password: megan/megan and start an experiment, then I would need to log back in with the same username and password to see those experiments. The username and password, however, does not limit your access to Driverless AI. If you want to use a new user name and password, you can log in again with a new username and password, but keep in mind that you won’t see your old experiments.

Can Driverless AI run on CPU-only machines?

Yes, Driverless AI can run on machines with CPUs only, though GPUs are recommended. Installation instructions are available for GPU and CPU systems. Refer to Installing and Upgrading Driverless AI for more information.

How often do new versions come out?

Driverless AI is in early stage development, and new releases come out quite frequently. Currently, you can expect a new release approximately every 7 to 10 days.

How can I upgrade to a newer version of Driverless AI?

Upgrade instructions vary depending on your environment. Refer to the installation section for your environment. Upgrade instructions are included there.

What kind of authentication is supported in Driverless AI?

Driverless AI supports LDAP, PAM, and Local authentication. These can be configured by setting the appropriate environment variables in the config.toml file or by specifying the environment variables when starting Driverless AI. Refer to Configuring Authentication for more information.

Can I set up SSL on Driverless AI?

Yes, you can set up HTTPS/SSL on Driverless AI running in an AWS environment. HTTPS/SSL needs to be configured on the host machine, and the necessary ports will need to be opened on the AWS side. You will need to have your own SSL cert or you can create a self signed cert for yourself.

The following is a very simple example showing how to configure HTTPS with a proxy pass to the port on the container 12345 with the keys placed in /etc/nginx/. Replace <server_name> with your server name.

server {
    listen 80;
    return 301 https://$host$request_uri;

server {
    listen 443;

    # Specify your server name here
    server_name <server_name>;

    ssl_certificate           /etc/nginx/cert.crt;
    ssl_certificate_key       /etc/nginx/cert.key;
    ssl on;
    ssl_session_cache  builtin:1000  shared:SSL:10m;
    ssl_protocols  TLSv1 TLSv1.1 TLSv1.2;
    ssl_ciphers HIGH:!aNULL:!eNULL:!EXPORT:!CAMELLIA:!DES:!MD5:!PSK:!RC4;
    ssl_prefer_server_ciphers on;

    access_log            /var/log/nginx/dai.access.log;

    location / {
      proxy_set_header        Host $host;
      proxy_set_header        X-Real-IP $remote_addr;
      proxy_set_header        X-Forwarded-For $proxy_add_x_forwarded_for;
      proxy_set_header        X-Forwarded-Proto $scheme;

      # Fix the “It appears that your reverse proxy set up is broken" error.
      proxy_pass          http://localhost:12345;
      proxy_read_timeout  90;

      # Specify your server name for the redirect
      proxy_redirect      http://localhost:12345 https://<server_name>;

More information about SSL for Nginx in Ubuntu 16.04 can be found here: https://www.digitalocean.com/community/tutorials/how-to-create-a-self-signed-ssl-certificate-for-nginx-in-ubuntu-16-04.


Is there a file size limit for datasets?

For GBMs, the file size for datasets is limited by the collective CPU or GPU memory on the system, but we continue to make optimizations for getting more data into an experiment, such as using TensorFlow streaming to stream to arbitrarily large datasets.

How much memory does Driverless AI require in order to run experiments?

Right now, Driverless AI requires approximately 10x the size of the data in system memory.

How many columns can Driverless AI handle?

Driverless AI has been tested on datasets with 10k columns. When running experiments on wide data, Driverless AI automatically checks if it is running out of memory, and if it is, it reduces the number of features until it can fit in memory. This may lead to a worse model, but Driverless AI shouldn’t crash because the data is wide.

How should I use Driverless AI if I have large data?

Driverless AI can handle large datasets out of the box. For very large datasets (more than 10 billion rows x columns), we recommend sampling your data for Driverless AI. Keep in mind that the goal of driverless AI is to go through many features and models to find the best modeling pipeline, and not to just train a few models on the raw data (H2O-3 is ideally suited for that case).

For large datasets, the recommended steps are:

  1. Run with the recommended accuracy/time/interpretability settings first, especially accuracy <= 7
  2. Gradually increase accuracy settings to 7 and choose accuracy 9 or 10 only after observing runs with <= 7.

How does Driverless AI detect the ID column?

The ID column logic is that the column is named ‘id’, ‘Id’, ‘ID’ or ‘iD’ exactly, or whether it contains unique integers.

Can Driverless AI handle data with missing values/nulls?

Yes, data that is imported into Driverless AI can include missing values. Feature engineering is fully aware of missing values, and missing values are treated as information - either as a special categorical level or as a special number. So for target encoding, for example, rows with a certain missing feature will belong to the same group. For Categorical Encoding where aggregations of a numeric columns are calculated for a grouped categorical column, missing values are kept. The formula for calculating the mean is the sum of non-missing values divided by the count of all values. For clustering, we impute missing values. And for frequency encoding, we count the number of rows that have a certain missing feature.

How are outliers handled?

Outliers are not removed from the data. Instead Driverless AI finds the best way to represent data with outliers. For example, Driverless AI may find that binning a variable with outliers improves performance.

For target columns, Driverless AI first determines the best representation of the column. It may find that for a target column with outliers, it is best to predict the log of the column.

If I drop several columns from the Train dataset, will Driverless AI understand that it needs to drop the same columns from the Test dataset?

If you drop columns from the training dataset, Driverless AI will do the same for the validation and test datasets (if the columns are present). There is no need for these columns because no features will be created from them.

Does Driverless AI treat numeric variables as categorical variables?

In certain cases, yes. You can prevent this behavior by setting the num_as_cat variable in your installation’s config.toml file to false. You can have finer grain control over this behavior by blacklisting the Numeric to Categorical Target Encoding Transformer and the Numeric To Categorical Weight of Evidence Transformer and their corresponding genes in your installation’s config.toml file. To learn more about the config.toml file, see the Using the config.toml File section.

Which algorithms are used in Driverless AI?

Features are engineered with a proprietary stack of Kaggle-winning statistical approaches including some of the most sophisticated target encoding and likelihood estimates based on groupings, aggregations and joins, but we also employ linear models, neural nets, clustering and dimensionality reduction models and many traditional approaches such as one-hot encoding etc.

On top of the engineered features, sophisticated models are fitted, including, but not limited to: XGBoost (both original XGBoost and ‘lossguide’ (LightGBM) mode), GLM if enabled via config.toml, TensorFlow (including a TensorFlow NLP recipe based on CNN Deeplearning models), and RuleFit. More models will be added later.

In general, GBMs are the best single-shot algorithms. Since 2006, boosting methods have proven to be the most accurate for noisy predictive modeling tasks outside of pattern recognition in images and sound (https://www.cs.cornell.edu/~caruana/ctp/ct.papers/caruana.icml06.pdf). The advent of XGBoost and Kaggle only cemented this position.

What objective function is used in XGBoost?

The objective function used in XGBoost is:

  • reg:linear and custom absolute error objective function for regression
  • binary:logistic or multi:softprob for classification

The objective function does not change depending on the scorer chosen. The scorer influences parameter tuning only.

For regression, Tweedie/Gamma/Poisson/etc. regression is not yet supported, but Driverless AI handles various target transforms so many target distributions can be handled very effectively already. Driverless AI handles quantile regression for alpha=0.5 (media), and general quantiles are on the roadmap.

Further details for the XGBoost instantiations can be found in the logs and in the model summary, both of which can be downloaded from the GUI or are found in the /tmp/h2oai_experiment_<name>/ folder on the server.

Does Driverless AI perform internal or external validation?

Driverless AI does internal validation when only training data is provided. It does external validation when training and validation data are provided. In either scenario, the validation data is used for all parameter tuning (models and features), not just for feature selection. Parameter tuning includes target transformation, model selection, feature engineering, feature selection, stacking, etc.


  • Internal validation (only training data given):
    • Ideal when data is either close to i.i.d., or for time-series problems
    • Internal holdouts are used for parameter tuning, with temporal causality for time-series problems
    • Will do the full spectrum from single holdout split to 5-fold CV, depending on accuracy settings
    • No need to split training data manually
    • Final models are trained using CV on the training data
  • External validation (training + validation data given):
    • Ideal when there’s some amount of drift in the data, and the validation set mimics the test set data better than the training data
    • No training data wasted during training because training data not used for parameter tuning
    • Validation data is used only for parameter tuning, and is not part of training data
    • No CV possible because we explicitly do not want to overfit on the training data
    • Not allowed for time-series problems (see below)

Tip: If you want both training and validation data to be used for parameter tuning (the training process), just concatenate the datasets together and turn them both into training data for the “internal validation” method.

How does Driverless AI prevent overfitting?

Driverless AI performs a number of checks to prevent overfitting. For example, during certain transformations, Driverless AI calculates the average on out-of-fold data using cross validation. Driverless AI also performs early stopping for every model built, ensuring that the model build will stop when it ceases to improve on holdout data. And additional steps to prevent overfitting include checking for i.i.d. and avoiding leakage during feature engineering.

A blog post describing Driverless AI overfitting protection in greater detail is available here: https://blog.h2o.ai/2018/03/driverless-ai-prevents-overfitting-leakage/.

How’s does Driverless AI avoid the multiple hypothesis (MH) problem?

Or more specifically, like many brute force methods for tuning hyperparameters/model selection, Driverless AI runs up against the multihypothesis problem (MH). For example, if I randomly generated a gazillion models, the odds that a few will do awesome on the test data that they are all measured against is pretty large, simply by sheer luck. How does Driverless AI address this?

Driverless AI uses a variant of the reusable holdout technique to address the multiple hypothesis problem. Refer to https://pdfs.semanticscholar.org/25fe/96591144f4af3d8f8f79c95b37f415e5bb75.pdf for more information.

How does Driverless AI suggest the experiment settings?

When you run an experiment on a dataset, the experiment settings (Accuracy, Time, and Interpretability) are automatically suggested by Driverless AI. For example, Driverless AI may suggest the parameters Accuracy = 7, Time = 3, Interpretability = 6, based on your data.

Driverless AI will automatically suggest experiment settings based on the number of columns and number of rows in your dataset. The settings are suggested to ensure best handling when the data is small. If the data is small, Driverless AI will suggest the settings that prevent overfitting and ensure the full dataset is utilized.

If the number of rows and number of columns are each below a certain threshold, then:

  • Accuracy will be increased up to 8.
    • The accuracy is increased so that cross validation is done. (We don’t want to “throw away” any data for internal validation purposes.)
  • Interpretability will be increased up to 8.
    • The higher the interpretability setting, the smaller the number of features in the final model.
    • More complex features are not allowed.
    • This prevents overfitting.
  • Time will be decreased down to 2.
    • There will be fewer feature engineering iterations to prevent overfitting.

What happens when I set Interpretability and Accuracy to the same number?

The answer is currently that interpretability controls which features are created and what features are kept. (Also above interpretability = 6, monotonicity constraints are used in XGBoost.) The accuracy refers to how hard Driverless AI then tries to make those features into the most accurate model.

Can I specify the number of GPUs to use when running Driverless AI?

When running an experiment, the Expert Settings allow you to specify the starting GPU ID for Driverless AI to use. You can also specify the maximum number of GPUs to use per model and per experiment. Refer to the Expert Settings section for more information.

Why is my experiment suddenly slow?

It is possible that your experiment has gone from using GPUs to using CPUs due to a change of the host system outside of Driverless AI’s control. You can verify this using any of the following methods:

  • Check GPU usage by going to your Driverless AI experiment page and clicking on the GPU USAGE tab in the lower-right quadrant of the experiment.
  • Run nvidia-smi in a terminal to see if any processes are using GPU resources in an unexpected way (such as those using a large amount of memory).
  • Check if System/GPU memory is being consumed by prior jobs or other tasks or if older jobs are still running some tasks.
  • Check and diable automatic NVIDIA driver updates on your system (as they can interfere with running experiments).

The general solution to these kind of sudden slowdown problems is to restart:

  • Restart Docker if using Docker
  • pkill --signal 9 h2oai if using the native installation method
  • Restart the system if nvidia-smi does not work as expected (e.g., after a driver update)

More ML-related issues that can lead to a slow experiment are:

  • Choosing high accuracy settings on a system with insufficient memory
  • Choosing low interpretability settings (can lead to more feature engineering which can increase memory usage)
  • Using a dataset with a lot of columns (> 500)
  • Doing multi-class classification with a GBM model when there are many target classes (> 5)

Why does the final model performance appear to be worse than previous iterations?

There are a few things to remember:

  • Driverless AI creates a best effort estimate of the generalization performance of the best modeling pipeline found so far
  • The performance estimation is always based on holdout data (data unseen by the model).
  • If no validation dataset is provided, the training data is split internally to create internal validation holdout data (once or multiple times or cross-validation, depending on the accuracy settings).
  • If no validation dataset is provided, for accuracy <= 7, a single holdout split is used, and a “lucky” or “unlucky” split can bias estimates for small datasets or datasets with high variance.
  • If a validation dataset is provided, then all performance estimates are solely based on the entire validation dataset (independent of accuracy settings).
  • All scores reported are based on bootstrapped-based statistical methods and come with error bars that represent a range of estimate uncertainty.

After the final iteration, a best final model is trained on a final set of engineered features. Depending on accuracy settings, a more accurate estimation of generalization performance may be done using cross-validation. Also, the final model may be a stacked ensemble consisting of multiple base models, which generally leads to better performance. Consequently, in rare cases, the difference in performance estimation method can lead to the final model’s estimated performance seeming poorer than those from previous iterations. (i.e., The final model’s estimated score is significantly worse than the last iteration score and error bars don’t overlap.) In that case, it is very likely that the final model performance estimation is more accurate, and the prior estimates were biased due to a “lucky” split. To confirm this, you can re-run the experiment multiple times (without setting the reproducible flag).

If you would like to minimize the likelihood of the final model performance appearing worse than previous iterations, here are some recommendations:

  • Increase accuracy settings
  • Provide a validation dataset
  • Provide more data

How can I find features that may be causing data leakages in my Driverless AI model?

To find original features that are causing leakage, have a look at features_orig.txt in the experiment summary download. Features causing leakage will have high importance there. To get a hint at derived features that might be causing leakage, create a new experiment with dials set to 2/2/8, and run the new experiment on your data with all your features and response. Then analyze the top 1-2 features in the model variable importance. They are likely the main contributors to data leakage if it is occurring.

How can I see the performance metrics on the test data?

As long as you provide a target column in the test set, Driverless AI will show the best estimate of the final model’s performance on the test set at the end of the experiment. The test set is never used to tune parameters (unlike to what Kagglers often do), so this is purely a convenience. Of course, you can still make test set predictions and compute your own metrics using a method of your choice.

How can I see all the performance metrics possible for my experiment?

At the end of the experiment, the model’s estimated performance on all provided datasets with a target column is printed in the experiment logs. For example, for the test set:

Final scores on test (external holdout) +/- stddev:
               GINI = 0.87794 +/- 0.035305 (more is better)
                MCC = 0.71124 +/- 0.043232 (more is better)
                F05 = 0.79175 +/- 0.04209 (more is better)
                 F1 = 0.75823 +/- 0.038675 (more is better)
                 F2 = 0.82752 +/- 0.03604 (more is better)
           ACCURACY = 0.91513 +/- 0.011975 (more is better)
            LOGLOSS = 0.28429 +/- 0.016682 (less is better)
              AUCPR = 0.79074 +/- 0.046223 (more is better)
 optimized: AUC = 0.93386 +/- 0.018856 (more is better)

What if my training/validation and testing data sets come from different distributions?

In general, Driverless AI uses training data to engineer features and train models and validation data to tune all parameters. If no external validation data is given, the training data is used to create internal holdouts. The way holdouts are created internally depends on whether there is a strong time dependence, see the point below. If the data has no obvious time dependency (e.g., if there is no time column neither implicit or explicit), or if the data can be sorted arbitrarily and it won’t affect the outcome (e.g., Iris data, predicting flower species from measurements), and if the test dataset is different (e.g., new flowers or only large flowers), then the model performance on validation (either internal or external) as measured during training won’t be achieved during final testing due to the obvious inability of the model to generalize.

What if my data has a time dependency?

If you know that your data has a strong time dependency, select a time column before starting the experiment. The time column must be in a Datetime format that can be parsed by pandas, such as “2017-11-06 14:32:21”, “Monday, June 18, 2012” or “Jun 18 2018 14:34:00” etc., or contain only integers.

If you are unsure about the strength of the time dependency, run two experiments: One with time column set to “[OFF]” and one with time column set to “[AUTO]” (or pick a time column yourself).

Why can’t I specify a validation data set for time-series problems? Why do you look at the test set for time-series problems

The problem with validation vs test in the time series setting is that there is only one valid way to define the split. If a test set is given, its length in time defines the validation split and the validation data has to be part of train. Otherwise the time-series validation won’t be useful.

For instance: Let’s assume we have train = [1,2,3,4,5,6,7,8,9,10] and test = [12,13], where integers define time periods (e.g., weeks). For this example, the most natural train/valid split that mimics the test scenario would be: train = [1,2,3,4,5,6,7] and valid = [9,10], and month 8 is skipped for training. Note that we will look at the start time and the duration of the test set only (if provided), and not at the contents of the test data (neither features nor target). If the user provides validation = [8,9,10] instead of test data, then this could lead to inferior validation strategy and worse generalization. Hence, we use the user-given test set only to create the optimal internal train/validation splits. If no test set is provided, the user can provide the length of the test set (in periods), the length of the train/test gap (in periods) and the length of the period itself (in seconds).

Does Driverless AI handle weighted data?

Yes. You can optionally provide an extra weight column in your training (and validation) data with non-negative observation weights. This can be useful to implement domain-specific effects such as exponential weighting in time or class weights. All of our algorithms and metrics in Driverless AI support observation weights, but note that estimated likelihoods can be skewed as a consequence.

How does Driverless AI handle fold assignments for weighted data?

Currently, Driverless AI does not take the weights into account during fold creation, but you can provide a fold column to enforce your own grouping, i.e., to keep rows that belong to the same group together (either in train or valid). The fold column has to be a categorical column (integers ok) that assigns a group ID to each row. (It needs to have at least 5 groups because we do up to 5-fold CV.)

Where can I get details of the various transformations performed in an experiment?

Download the experiment’s log .zip file from the GUI. This zip file includes summary information, log information, and a gene_summary.txt file with details of the transformations used in the experiment. Specificially, there is a details folder with all subprocess logs.

On the server, the experiment specific files are inside the /tmp/h2oai_experiment_<name>/ folder after the experiment completes, particularly h2oai_experiment_logs_<name>.zip and h2oai_experiment_summary_<name>.zip.

How can I download the predictions onto the machine where Driverless AI is running?

When you select Score on Another Dataset, the predictions will automatically be stored on the machine where Driverless AI is running. They will be saved in the following locations (and can be opened again by Driverless AI, both for .csv and .bin):

  • Training Data Predictions: tmp/h2oai_experiment_<name>/train_preds.csv (also saved as .bin)
  • Testing Data Predictions: tmp/h2oai_experiment_<name>/test_preds.csv (also saved as .bin)
  • New Data Predictions: tmp/h2oai_experiment_<name>/automatically_generated_name.csv. Note that the automatically generated name will match the name of the file downloaded to your local computer.

Time Series

Why does the gap between train and test matter? Is it because of creating the lag features on the test set?

Taking the gap into account is necessary in order to avoid too optimistic estimates of the true error and to avoid creating history-based features like lags for the training and validation data (which cannot be created for the test data due to the missing information).

In regards to applying the target lags to different subsets of the time group columns, are you saying Driverless AI perform auto-correlation at “levels” of the time series? For example, consider the Walmart dataset where I have Store and Dept (and my target is Weekly Sales). Are you saying that Driverless AI checks for auto-correlation in Weekly Sales based on just Store, just Dept, and both Store and Dept?

Currently, auto-correlation is only applied on the detected superkey (entire TGC) of the training dataset relation at the very beginning. It’s used to rank potential lag-sizes, with the goal to prune the search space for the GA optimization process, which is responsible for selecting the lag features.

How does Driverless AI detect the time period?

Driverless AI treats each time series as a function with some frequency 1/ns. The actual value is estimated by the median of time deltas across maximal length TGC subgroups. The choosen SI unit minimizes the distance to all available SI units.

What is the logic behind the selectable numbers for forcast horizon length?

The shown forcast horizon options are based on quantiles of valid splits. This is necessary because Driverless AI cannot display all possible options in general.

Assume that in my Walmart dataset, all stores provided data at the week level, but one store provided data at the day level. What would Driverless AI do?

Driverless AI would still assume “weekly data” in this case because the majority of stores are yielding this property. The “daily” store would be resampled to the detected overall frequency.

Assume that in my Walmart dataset, all stores and departments provided data at the weekly level, but one department in a specific store provided weekly sales on a bi-weekly basis (every two weeks). What would Driverless AI do?

That’s similar to having missing data. Due to proper resampling, Driverless AI can handle this without any issues.

Why does the number of weeks that you want to start predicting matter?

That’s an option to provide a train-test gap if there is no test data is available. That is to say, “I don’t have my test data yet, but I know it will have a gap to train of x.”

Why isn’t it possible to provide a validation set?

The user cannot provide a validation set because, as opposed to the i.i.d. case, the test data (if given) already defines the validation data as part of train.

Are the scoring components of time series sensitive to the order in which new pieces of data arrive? I.e., is each row independent at scoring time, or is there a real-time windowing effect in the scoring pieces?

Each row is independent at scoring time.

What happens if the user, at predict time, gives a row with a time value that is too small or too large?

Internally, “out-of bounds” time values are encoded with special values. The samples will still be scored, but the predictions won’t be trustworthy.

What’s the minimum data size for a time series recipe?

We recommended that you have around 10,000 validation samples in order to get a reliable estimate of the true error. The time series recipe can still be applied for smaller data, but the validation error might be inaccurate.

How long must the training data be compared to the test data?

At a minimum, the training data has to be at least twice as long as the test data along the time axis. However, we recommended that the training data is at least three times as long as the test data.

How does the time series recipe deal with missing values?

Missing values will be converted to a special value, which is different from any non-missing feature value. Explicit imputation techniques won’t be applied.

Can the time information be distributed acrosss multiple columns in the input data (such as [year, day, month]?

Currently Driverless AI requires the data to have the time stamps given in a single column. Driverless AI will create addidtion time features like [year, day, month] on its own, if they turn out to be useful.

What type of modeling approach does Driverless AI use for time series?

Driverless AI combines the creation of history-based features like lags, moving averages etc. with the modeling techniques, which are also applied for i.i.d. data. The primary model of choice is XGBoost.

What is a lag, and why does it help?

A lag is a feature value from a previous point in time. Lags are useful to take advantage of the fact that the current (unknown) target value is often correlated with previous (known) target values. Hence, they can better capture target patterns along the time axis.

What’s the idea behind exponential weighting of moving averages?

Exponential weighting accounts for the possibility that more recent observatons are better suited to explain the present than older observations.