Depending on the goals we wish to achieve and the criteria by which we measure success, there are various methods to create artificial intelligence. The scope is extremely broad, ranging from highly complex systems like autonomous driving and robotics to common components of our daily lives, such as facial recognition, machine translation, and email classification, all of which fall under the domain of artificial intelligence.
After reading the article below, you may understand what is truly needed to create artificial intelligence.
“Do You Really Know What Artificial Intelligence Is?”
The path you choose will depend on what your AI’s goals are and your understanding of the complexity and feasibility of various methods. In this article, we will discuss those methods that are considered more feasible and universal for scientific development, namely the study of rational/intelligent agent design.
What Is an Intelligent Agent?
Perceives its environment through sensors
Acts upon it through actuators
It will operate in a cycle of perception, thinking, and action. For example, humans perceive the environment through their five senses (sensors), then think about it, and subsequently use their body parts (actuators) to perform actions. Similarly, machine intelligent agents perceive the environment through the sensors we provide (which can be cameras, microphones, infrared detectors), then perform some calculations (thinking), and finally use various motors/actuators to execute actions. Now, you should be clear that the world around you is filled with various intelligent agents, such as your phone, vacuum cleaner, smart refrigerator, thermostat, camera, and even yourself.
What Is an Intelligent Agent?
An Intelligent Agent is a type of agent that operates in a way that is expected to maximize its performance metrics, given what it perceives and its prior knowledge.
Performance metrics define the criteria for the agent’s success.
Such agents are also known as rational agents. The rationality of an agent is measured by its performance metrics, the prior knowledge it possesses, the environment it can perceive, and the actions it can perform.
This concept is central to artificial intelligence.
The aforementioned properties of Intelligent Agents are often summarized by the term PEAS (Performance, Environment, Actuators, and Sensors), which represents Performance, Environment, Actuators, and Sensors. For instance, an autonomous vehicle should have the following PEAS:
Performance:Safety, time, legal driving, comfort.
Environment:Roads, other vehicles, pedestrians, traffic signs.
Actuators:Steering, accelerator, brakes, signals, horn.
Sensors:Cameras, sonar, GPS, speedometer, odometer, accelerometer, engine sensors, keyboard.
To meet the use cases in the real world, artificial intelligence itself needs to have a wide range of Intelligent Agents. This introduces the diversity of the types of agents and environments we have. Next, let’s take a look together.
Types of Environments
If you want to design a rational agent, you must keep in mind the types of environments it will operate in, which can be categorized as follows:
Fully Observable and Partially Observable:
If fully observable, the agent’s sensors can access the complete state of the environment at every point in time; otherwise, they cannot. For example, chess is a fully observable environment, while poker is not.
Deterministic and Stochastic:
The next state of the environment is fully determined by the current state and the actions the agent takes next. (If the environment is deterministic, but the behavior of other agents is uncertain, then the environment is stochastic). Stochastic environments are inherently random and cannot be fully determined. For example, the 8-puzzle game has a deterministic environment, but a self-driving car does not.
Static and Dynamic:
A static environment does not change while the agent is deliberating. (A semi-dynamic environment does not change over time, but the agent’s performance score does change accordingly). On the other hand, a dynamic environment changes. Backgammon has a static environment, while a robot vacuum cleaner has a dynamic environment.
Discrete and Continuous:
A discrete environment consists of a finite number of well-defined perceptions and actions. For example, checkers is a discrete environment, while an autonomous vehicle needs to operate in a continuous environment.
Single-Agent and Multi-Agent:
An agent that only operates itself has a single-agent environment. However, if other agents are included, it becomes a multi-agent environment. Autonomous vehicles operate in a multi-agent environment.
There are also other types of environments, such as situational and sequential, known and unknown, which define the scope of agents.
Types of Agents
There are generally four types of agents, distinguished by their level of intelligence or the complexity of tasks they can perform. All types can improve performance over time and produce better actions. These can be summarized as learning agents.
Simple Reflex Agents
These agents select actions based solely on the current state, ignoring the history of perceptions.
They can only work if the environment is fully observable or if the correct behavior is based on the current perception.
Model-Based Reflex Agents
These agents track partially observable environments. Their internal states depend on the history of perceptions. The modeling of the environment/world is based on how it evolves independently of the agent and how the agent’s actions affect the world.
Goal-Based Agents
This is an improvement over model-based agents and is used when the current state of the environment is not sufficient. The agent combines the goal information provided with the environmental model to select actions that achieve that goal.
Utility-Based Agents
Improvements over goal-based agents are not sufficient for achieving expected goals. We may need to consider costs. For example, we may seek a faster, safer, and cheaper journey to reach our destination. This is marked by a utility function. Utility agents will choose actions that maximize expected utility.
General agents, also known as learning agents, proposed by Alan Turing, represent the most advanced approach in current artificial intelligence systems.
All of the above agents can be summarized as learning agents to produce better actions.
Learning Agents
Learning components: responsible for improvement
Performance components: responsible for selecting external behaviors, which is what we generally consider to be an agent.
Comment: How does the agent perform based on determined performance standards?
Problem generator: allows the agent to explore.
Internal State Representation
As agents become more complex, their internal structures also become more complex. The way they store internal states changes. Simple reflex agents do not need to store states due to their nature, while other types do. The image below provides a high-level representation of agent states, arranged in order of increasing representational capability (from left to right).
Atomic Representation:
In this case, the state is stored as a black box, meaning there is no internal structure. For example, for Roomba (robot vacuum cleaner), the internal state is the patch that has already been vacuumed, and you do not need to know anything else. As shown, this representation is suitable for model-based and goal-based agents and is used in various AI algorithms such as search problems and adversarial games.
Factored Representation:
In this representation, the state is no longer a black box. It now has attribute-value pairs, also known as variables that can contain values. For example, in finding a route, you have GPS location and fuel level in the tank. This adds a constraint to the problem. As shown, this representation is suitable for goal-based agents and is used in various AI algorithms such as constraint satisfaction and Bayesian networks.
Structured Representation:
In this representation, we have relationships between variable/factor states, which raises logic in AI algorithms. For example, in natural language processing, whether a statement contains a reference to someone and whether the adjectives in that statement describe that person. These relational statements will determine whether the statement is sarcastic. This is high-level artificial intelligence used in first-order logic, knowledge-based learning, and natural language understanding algorithms.
For artificial intelligence, there are many more rational agents; this is just an overview. As you know, the design study of rational agents is an important component of artificial intelligence, as it has widespread applications in various fields. However, these agents cannot work independently; they require an artificial intelligence algorithm to drive them. Most of these algorithms involve search. Soon, I will write some articles about AI algorithms that drive rational agents and the use of machine learning in AI.
Source: AI Scientist WeChat Official Account
