API Reference - Models
If you wonder what are the most high-value use cases that helps with retention and revenue generation this DataPredict™, you can view them here!
| Model Type | Count |
|---|---|
| Regression | 8 |
| Classification | 13 |
| Clustering | 8 |
| Deep Reinforcement Learning | 21 |
| Tabular Reinforcement Learning | 12 |
| Sequence Modelling | 6 |
| Generative | 4 |
| Total | 72 |
Legend
| Icon | Name | Description |
|---|---|---|
| ❗ | Implementation Issue | The model may have some implementation problems. |
| 🔰 | Beginner Algorithm | Commonly taught to beginners. |
| 💾 | Data Efficient | Require few data to train the model. |
| ⚡ | Computationally Efficient | Require few computational resources to train the model. |
| 🛡️ | Noise Resistant | Can handle randomness / unclean data. |
| 🟢 | Online | Can adapt real-time. |
| 🟡 | Session-Adaptive / Offline | Can be retrained each session. |
| ⚠️ | Assumption-Heavy | Assumes linear / independent features. |
Note
-
For strong deep learning applications, have a look at DataPredict™ Neural (object-oriented) and DataPredict™ Axon (function-oriented) instead. DataPredict™ is only suitable for general purpose machine, deep and reinforcement learning.
-
Contains most of the deep reinforcement learning and generative algorithms listed here.
-
Includes convolutional, pooling, embedding, dropout and activation layers.
-
Uses reverse-mode automatic differentiation and lazy differentiation evaluation for DataPredict™ Neural (static graph) and DataPredict™ Axon (dynamic graph).
-
-
Currently, DataPredict™ has ~90% (61 out of 70) models with online learning capabilities. By default, most models would perform offline / batch training on the first train, but then switches to online / incremental / sequential after the first train.
-
Tabular reinforcement learning models can use optimizers. And yes, I am quite aware that I have overengineered this, but I really want to make this a grand finale before I stop updating DataPredict™ for a long time.
-
No dimensionality reduction algorithms due to not being suitable for game-related use cases. They tend to be computationally expensive and are only useful when a full dataset is collected. This can be offset by choosing proper features and remove the unnecessary ones.
-
Going “Gold” on my birthday at 23 January 2026. Probably.
Regression
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| LinearRegression | LR | 🔰 🟢 🟡 | General Time-To-Leave Prediction And In-Game Currency Price Generation |
| QuantileLinearRegression | None | 🟢 🟡 | Case-Based Time-To-Leave Prediction And In-Game Currency Price Generation |
| PassiveAggressiveRegressor | PA-R | ⚡ 🟢 | Fast Constrained Time-To-Leave Prediction And In-Game Currency Price Generation |
| SupportVectorRegression | SVR | 💾 🟡 | Constrained Time-To-Leave Prediction And In-Game Currency Price Generation |
| KNearestNeighboursRegressor | KNN-R | 🟢 🟡 | Memory-Based Time-To-Leave Prediction And In-Game Currency Price Generation |
| NormalLinearRegression* | None | 💾 ⚡ 🟡 ⚠️ | Instant Train Time-To-Leave Prediction And In-Game Currency Price Generation |
| BayesianLinearRegression* | None | 💾 ⚡ 🟡 ⚠️ | Instant Train Time-To-Leave Prediction And In-Game Currency Price Generation With Probability Estimation |
| BayesianQuantileLinearRegression* | None | 💾 ⚡ 🟡 ⚠️ | Instant Train Time-To-Leave Prediction And In-Game Currency Price Generation With Case Estimation |
* The “instant train” models have these issues:
-
It assumes that the features have a linear relationship with the label values, which is almost certainly not true in game-related settings. Hence, it is recommended to add small independent noise values to each features.
-
The feature matrix will also need to have shape of (n x n). This naturally leads to the requirement of label vector with a shape of (n x 1).
Classification
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| LogisticRegression | Perceptron, Sigmoid Regression | 🔰 🟢 🟡 | Probability-To-Leave Prediction, Player Churn Prediction, Confidence Prediction |
| PassiveAggressiveClassifier | PA-C | ⚡ 🟢 | Fast Purchase Likelihood Estimation, Decision Making |
| OneClassPassiveAggressiveClassifier | OC-PA-C | ⚡ 🟢 | Fast Hacking Detection, Anomaly Detection (Using Single Class Data) |
| NearestCentroid | NC | ⚡ 🟢 🟡 | Fast Grouping Or Quick Decision Making |
| KNearestNeighboursClassifier | KNN-C | 🟢 🟡 | Item Recommendation, Similar Player Matchmaking |
| SupportVectorMachine | SVM | 💾 🟡 | Hacking Detection, Anomaly Detection |
| OneClassSupportVectorMachine | OC-SVM | 💾 🟡 | Hacking Detection, Anomaly Detection (Using Single Class Data) |
| NeuralNetwork | Multi-Layer Perceptron | 🟢 🟡 | Decision-Making, Player Behaviour Prediction |
| GaussianNaiveBayes* | GNB | 💾 ⚡ 🟢 🟡 ⚠️ | Enemy Data Generation, Player Behavior Categorization (e.g. Cautious Vs. Aggressive), Fast State Classification |
| MultinomialNaiveBayes* | MNB | 💾 ⚡ 🟢 🟡 ⚠️ | Summoning Next Enemy Type, Inventory Action Prediction, Strategy Profiling Based on Item Usage |
| BernoulliNaiveBayes* | BNB | 💾 ⚡ 🟢 🟡 ⚠️ | Binary Action Prediction (e.g. Jump Or Not), Quick Decision Filters |
| ComplementNaiveBayes* | CNB | 💾 ⚡ 🟢 🟡 ⚠️ | Imbalanced Class Prediction (e.g. Rare Choices, Niche Paths) |
| CategoricalNaiveBayes* | CNB | 💾 ⚡ 🟢 🟡 ⚠️ | Player Choice Prediction (e.g. Weapon Type, Character Class, Map Region Selection) |
* “Naive Bayes” models assumes that the features are independent to each other, which is almost certainly not true in game-related settings. Additionally, these models are better as generative models, despite being commonly taught as a classifier.
Clustering
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| KMeans | None | 🔰 🟢 🟡 | Maximizing Area-of-Effect Abilities, Target Grouping |
| FuzzyCMeans | None | 🟢 🟡 | Overlapping Area-of-Effect Abilities, Overlapping Target Grouping |
| KMedoids | None | 🟢 🟡 | Player Grouping Based On Player Locations With Leader Identification |
| AgglomerativeHierarchical | None | 🟢 🟡 | Enemy Data Generation |
| ExpectationMaximization | EM | 🟢 🟡 | Hacking Detection, Anomaly Detection |
| MeanShift | None | 🛡️ 🟢 🟡 | Boss Spawn Location Search Based On Player Locations |
| AffinityPropagation | AP | 🟡 | Player Grouping |
| DensityBasedSpatialClusteringOfApplicationsWithNoise | DBSCAN | 🛡️ 🟡 | Density Grouping |
Deep Reinforcement Learning
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| DeepQLearning | Deep Q Network | 💾 🟢 | Best Self-Learning Player AIs, Best Recommendation Systems |
| DeepDoubleQLearningV1 | Double Deep Q Network (2010) | 💾 🛡️ 🟢 | Stable Best Self-Learning Player AIs, Best Recommendation Systems |
| DeepDoubleQLearningV2 | Double Deep Q Network (2015) | 💾 🛡️ 🟢 | Stable Best Self-Learning Player AIs, Best Recommendation Systems |
| DeepClippedDoubleQLearning | Clipped Deep Double Q Network | 💾 🛡️ 🟢 | Stable Best Self-Learning Player AIs, Best Recommendation Systems |
| DeepStateActionRewardStateAction | Deep SARSA | 🟢 | Safe Self-Learning Player AIs, Safe Recommendation Systems |
| DeepDoubleStateActionRewardStateActionV1 | Double Deep SARSA | 🛡️ 🟢 | Stable Safe Self-Learning Player AIs, Safe Recommendation Systems |
| DeepDoubleStateActionRewardStateActionV2 | Double Deep SARSA | 🛡️ 🟢 | Stable Safe Self-Learning Player AIs, Safe Recommendation Systems |
| DeepExpectedStateActionRewardStateAction | Deep Expected SARSA | 🟢 | Balanced Self-Learning Player AIs, Balanced Recommendation Systems |
| DeepDoubleExpectedStateActionRewardStateActionV1 | Double Deep Expected SARSA | 🛡️ 🟢 | Stable Balanced Self-Learning Player AIs, Balanced Recommendation Systems |
| DeepDoubleExpectedStateActionRewardStateActionV2 | Double Deep Expected SARSA | 🛡️ 🟢 | Stable Balanced Self-Learning Player AIs, Balanced Recommendation Systems |
| DeepMonteCarloControl | None | ❗ 🟢 | Online Self-Learning Player AIs |
| DeepOffPolicyMonteCarloControl | None | 🟢 | Offline Self-Learning Player AIs |
| REINFORCE | None | 🟢 | Reward-Based Self-Learning Player AIs |
| VanillaPolicyGradient | VPG | ❗ 🟢 | Baseline-Based Self-Learning Player AIs |
| ActorCritic | AC | 🟢 | Critic-Based Self-Learning Player AIs |
| AdvantageActorCritic | A2C | 🟢 | Advantage-Based Self-Learning Player AIs |
| ProximalPolicyOptimization | PPO | 🟢 | Industry-Grade And Research-Grade Self-Learning Player And Vehicle AIs |
| ProximalPolicyOptimizationClip | PPO-Clip | 🟢 | Industry-Grade And Research-Grade Self-Learning Player And Vehicle AIs |
| SoftActorCritic | SAC | 💾 🛡️ 🟢 | Self-Learning Vehicle AIs |
| DeepDeterministicPolicyGradient | DDPG | 🟢 | Self-Learning Vehicle AIs |
| TwinDelayedDeepDeterministicPolicyGradient | TD3 | 🟢 🛡️ | Self-Learning Vehicle AIs |
Tabular Reinforcement Learning
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| TabularQLearning | Q-Learning | 🔰 💾 🟢 | Best Self-Learning Grid AIs |
| TabularDoubleQLearningV1 | Double Q-Learning (2010) | 💾 🛡️ 🟢 | Best Self-Learning Grid AIs |
| TabularDoubleQLearningV2 | Double Q-Learning (2015) | 💾 🛡️ 🟢 | Best Self-Learning Grid AIs |
| TabularClippedDoubleQLearning | Clipped Double Q-Learning | 💾 🛡️ 🟢 | Best Self-Learning Grid AIs |
| TabularStateActionRewardStateAction | SARSA | ❗ 🔰 🟢 | Safe Self-Learning Grid AIs |
| TabularDoubleStateActionRewardStateActionV1 | Double SARSA | ❗ 🛡️ 🟢 | Safe Self-Learning Grid AIs |
| TabularDoubleStateActionRewardStateActionV2 | Double SARSA | ❗ 🛡️ 🟢 | Safe Self-Learning Grid AIs |
| TabularExpectedStateActionRewardStateAction | Expected SARSA | 🟢 | Balanced Self-Learning Grid AIs |
| TabularDoubleExpectedStateActionRewardStateActionV1 | Double Expected SARSA | 🛡️ 🟢 | Balanced Self-Learning Grid AIs |
| TabularDoubleExpectedStateActionRewardStateActionV2 | Double Expected SARSA | 🛡️ 🟢 | Balanced Self-Learning Grid AIs |
| TabularMonteCarloControl | MC | 🟢 | Online Self-Learning Grid AIs |
| TabularOffPolicyMonteCarloControl | Off-Policy MC | 🟢 | Offline Self-Learning Grid AIs |
Sequence Modelling
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| Markov* | None | 💾 🟢 | Single Features Player State Prediction |
| Dynamic Bayesian Network* | DBN | 💾 🟢 | Multiple Features Player State Prediction |
| Conditional Random Field* | CRF | 🟢 | Multiple Features Player State Prediction |
| Kalman Filter* | KF | 🟢 ⚠️ | Multiple Features Player State Prediction |
| Extended Kalman Filter* | EKF | ⚡ 🟢 | Multiple Features Player State Prediction |
| Unscented Kalman Filter* | UKF | ⚡ 💾 🟢 | Multiple Features Player State Prediction |
- These are single step variants of the sequence models. Hence, it will not use or return sequence of values.
Generative
❗Implementation Issue 🔰 Beginner Algorithm 💾 Data Efficient ⚡ Computationally Efficient 🛡️ Noise Resistant 🟢 Online 🟡 Session-Adaptive / Offline ⚠️ Assumption-Heavy
| Model | Alternate Names | Properties | Use Cases |
|---|---|---|---|
| GenerativeAdversarialNetwork | GAN | 🟢 🟡 | Enemy Data Generation |
| ConditionalGenerativeAdversarialNetwork | CGAN | 🟢 🟡 | Conditional Enemy Data Generation |
| WassersteinGenerativeAdversarialNetwork | WGAN | 🟢 🟡 | Stable Enemy Data Generation |
| ConditionalWassersteinGenerativeAdversarialNetwork | CWGAN | 🟢 🟡 | Stable Conditional Enemy Data Generation |
BaseModels
DeepReinforcementLearningBaseModel
DeepReinforcementLearningActorCriticBaseModel