**
3203. Introduction to Artificial Intelligence**
**Learning: Statistical Approaches**

### 1. Learning as statistical inference

Since statistics studies the analysis and interpretation of data, many existing statistical methods can be considered as learning methods:
- Bayes Theorem is often used to learn a posterior probability distribution from a prior probability distribution and a piece of evidence, as well as a classifier.
- Regression analysis can be considered as learning a general function from given values.
- Parameter estimation can be taken as a process of learning a probability distribution from background knowledge and empirical data.
- Clustering classifies data into groups according to their similarity, so can be considered as unsupervised concept learning. One example is the expectation–maximization (EM) algorithm.

Furthermore, many neural networks can be analyzed as implementing statistical models. Therefore, the current machine learning study has a significant overlap with statistics.

### 2. Learning belief networks

"Belief Network" is just another name of Bayesian network. Though this model has been widely used for reasoning with uncertainty, its design and construction is often a difficult task.
In certain situations, this task can be accomplished using certain machine leaning techniques, which can learn the involved probability values, as well as the network structure.

### 3. Reinforcement learning

In a general sense, "reinforcement learning" indicates the learning process in which the system's behaviors are changed by rewards and punishments obtained in the past.
An example is the Player 5, where the final result of a game is used to adjust preferences among options.

One hard problem in this type of learning is credit/blame assignment. When the feedback is only about a complete sequence of actions, not about each individual action (delayed reward), it is not always easy to determine what is right/wrong.

Another issue in reinforcement learning is the tradeoff between *exploration* and *exploitation*. To get the maximum reward in the long run in a uncertain environment, sometimes it is better to take a less-explored option, even when another option has a better historic record. One example is the
k-armed bandit problem.

In current machine learning study, the phrase
"reinforcement learning" refers to the problem of learning an optimal policy that maximizes the total expected reward in a Markov decision process. Consequently, the knowledge such a system has is statistical in nature.

### 4. Applications and limitations

Statistical machine learning is widely used in data mining and knowledge discovery, as well as in other responses toward the Big Data challenge.
Sometimes we find surprising relations in data:
diaper and beer,
football and election.
How to distinguish a coincidence from a causal relation?

Other major issues to be considered before applying statistical machine learning to a practical problem:

- Should the problem be considered as a statistical problem?
- Are there sufficient knowledge and data to satisfy the requirements of the methods involved?
- Are there sufficient computational resources to satisfy the requirements of the methods involved?