Choosing The Goal For The Model
In the general artificial intelligence space, there are multiple methods on how algorithms can learn from patterns.
| Method | Explanation |
|---|---|
| Supervised Learning | Learns patterns between the input features and the output features. |
| Unsupervised Learning | Learns patterns within the input features. |
| Semi-Supervised Learning | Use Supervised Learning predictions on inputs features that have unknown output features as its own training data. |
| Reinforcement Learning | Learns what actions to take for a set of input features that maximizes the likelihood of reaching a goal. |
- Note: Generative AIs falls under supervised learning because they require output features like text, images and audios. The generation part is the result of predicting from the input features.
Supervised Learning VS Reinforcement Learning
On the surface, these two may look like the same. However, the difference lies on how they are trained.
-
In supervised learning, we tell the algorithm that this input is what leads to that output.
-
In reinforcement learning, we don’t provide the output, but instead a score of how good the output is to reach a stated goal when the algorithm is given a particular input.
In other words, the supervised learning algorithms “translates” the input to a solution, while the reinforcement learning algorithms “searches” a solution for a given input.
What’s Your Goal?
-
To Achieve A Goal -> Use “measurement of engagement” metrics as rewards and combine it with reinforcement learning models.
-
Prediction -> Use regression and classification models.
-
Best Middle Values -> Use clustering models.
Supervised Learning VS Reinforcement Learning For Retention
Technically, you can use both supervised learning and reinforcement learning for retention. However, each of these methods have their own strengths and weakness.
-
Supervised learning will train faster than reinforcement learning as the output features provides an “anchor” for determining the hidden pattern between inputs and the outputs.
-
Reinforcement learning allows you to define an abstract measurement of fun (that are specific to individual players) into a single value and use that to find the best output for a given input.
In other words, the supervised learning tries to maximize the pattern-matching, while the reinforcement learning maximizes the likelihood of reaching the goal. They are complementary to each other in what they output.