I'm new to AI and and would like to understand the difference between learning agents and other types of agents, in what ways learning agents can be applied, and if they differ from deep learning.
The key difference between a learning agent and non-learning agents is that the learning agent can improve it's performance on it's own, allowing it to get "smarter".
Russel & Norvig cover the different types of intelligent agents in detail in their textbook Artificial Intelligence: A Modern Approach, and the wikipedia entry for intelligent agent mirrors their definitions. [Free pdf of the 3rd edition here.]
They break down agents into the following categories:
- Simple reflex agents
- Model-based reflex agents
- Goal-based agents
- Utility-based agents
- Learning agents
I'm not going to go into depth because understanding of technical terminology is required, but the links in the above list provide simple flow-chart explanations.
In the first 4, the more sophisticated agents contain the primitive agents. The model-based agent is actually a "model-based reflex agent", and the goal and utility agents are also model-based.
You will notice is that the structure changes fundamentally with the learning agent.
The key components of the learning agent are categorized as the "Critic", the "Learning Element", the "Performance Element", and "Problem Generator". This structure allows the agent to evaluate past performance and modify subsequent actions to improve performance (learning). The problem generator allows the agent to seek novel strategies (experimentation).
Where the simple reflex agent operates under a "condition-action rule" (if a condition is met, take this action), the learning agent can improve it's decision-making beyond it's initial programming.
Deep learning is a form of machine learning that uses representations. In the most general sense, the algorithms use a trial and error and statistical analysis to optimize performance. Various forms of neural networks are classified as deep learning systems. [See also Deep Reinforcement Learning vs. Reinforcement Learning.]
My sense is that learning agents can be applied to any type of problem in which parameters can be sufficiently defined. I don't want to get too specific because there are many types of learning algorithms, but we've been seeing application of machine learning (in various forms) expanding significantly since ~2016. This includes replacing classical control systems, medical diagnostics, natural language processing, image recognition, and consumer recommendations.
The applications are initially "narrow", defined as capability at a single task, but can be aggregated for tasks such as self-navigating vehicles which involve sets of discrete problems.
A learning agent is an agent which has a problem generator, a learning element, a critic and a performance element. ProbGen just creates different instances of a problem w.r.t. a domain and a set of parameters in jurisdiction. A learning element is responsible for making efficient choices and performance element chooses actions for the selected choice. It also changes some of the knowledge base. A critic is basically a metric to tell which is a better choice.
Deep learning is a bit complicated than that. Then again, let's simplify. You see, deep learning is a specific arrangement or subsection of neural networks that helps the learning agent learn faster. That's as simply as I could explain it. What we do is that we arrange the layers in a neural network with a purpose and densely pack it to make it more task specific. Obviously, I am leaving out a lot of implementation details but for them, I am afraid, you need to look up a tutorial.
How are they used in AI? That's problem specific. Learning agents are everywhere from spam detection to roombas to artificial dog pets. Deep learning on the other hand is currently focused on user pattern recognition and personalization in every front. Please let me know if you need me to expand upon this answer.