The ultimate goal of the development of technologies Artificial Intelligence (AI), led by companies like OpenAI, is to have artificial general intelligence (AGI) models, capable of solving any task or user request, and analyzing all types of information from different formats or sensory registers. And a key tool for building these general purpose models are the intelligent agents (or AI agents).
Intelligent agents are basically models developed for specific purposes, optimized to solve specific problems in different areas such as consumer assistance, internet of things (IoT) or in supply chains and logistics. Their purpose is to assist humans in relatively mechanical tasks, and by combining them, general-purpose models can be created.
A central characteristic of this technology is that its operation is autonomous. Although they are initially trained by humans, they can later operate independently, analyzing the information of a system or application or, in the case that they are applied to physical products such as robots, the environment that surrounds them. An example of use for this last case are cars with autonomous driving systems, which perceive this information through sensors and analyze it to define what actions to take.
In general terms there is two types of AI agents: those that follow predefined rules; and those who autonomously learn and adapt to different situations to respond in the best possible way.
In turn, these agents can be classified, depending on their functioning, as reagents (they respond directly to stimuli); deliberative (they plan to make different decisions); or with learning capabilities (they are adapted based on data and experience).
A use that reflects the possibilities offered by this type of agents is the one announced at the end of last year by the company OpenAI. After the rise of their conversational AI model, they presented a new tool for paying users, which allows them to create their own personalized agents according to the role needed, such as a “creative writing coach” or a “fellow traveler”. In this case, through natural language processing (NLP), you can interact with agents better prepared to solve specific tasks, compared to ChatGPT, which is general purpose.
Perceptions and actuators
For these models to come into operation, the first thing they must receive are the so-called “perceptions”, that is, the sensory inputs that the AI agent receives from its environment. These provide information about the current state of the observable environment in which the agent operates. For example, if it is a customer service chatbot, insights can include text messages, user profile information, their location, and even recognition of their emotions.
Another case would be that of a autonomous driving vehiclewhich receives information through its multiple sensors and processes it based on the parameters that make up its knowledge base.
Following the same example, the base could be composed of information about the roads, traffic rules such as speed limits and situations that were generated by humans during the training stage of the agents, before carrying out tests on the street.
Based on the information from the perceptions, the “actuators”, who are the ones who execute the actions based on the analysis. For example, an actuator can be a generator of text responses that are sent to the user in a support chat, or one that activates the brakes of a car if it detects a static object at a certain distance.
They can also be app notifications, such as a notification or alert email sent to a bank account holder when a transfer is made.
In addition to the work on these two elements, the feedback It is also essential for the improvement of AI agents over time. This feedback can come, on the one hand, from a human operator or from another AI system that supervises the agent in question.
On the other hand, the environment can also provide feedback in the form of results of the agent’s actions. This feedback loop allows the agent to adapt, learn from its experiences, and make better decisions in the future.
Types of AI agents
- Simple reflex agents: These agents, suitable for tasks with limited complexity, operate based on a set of predefined condition-action rules. They react to the current perception and do not consider the history of previous perceptions.
- Model-based reflection agents: model-based agents take a more advanced approach. They maintain an internal model of the environment and make decisions based on an understanding of their model.
- Utility-based agents: They make decisions considering the expected utility of each possible action. They are often used in situations where it is essential to weigh different options and select the one with the greatest expected utility.
- Learning agents: They are designed to operate in unknown environments. They learn from their experiences and adapt their actions over time. Deep learning and neural networks are often used in the development of learning agents.
- Belief-desire-intention agents: These agents shape human behavior by maintaining beliefs about the environment, desires, and intentions. They can reason and plan their actions accordingly, making them suitable for complex systems.
- Logic-based agents: They use deductive reasoning to make decisions, usually based on logical rules. They are well suited for tasks that require complex logical reasoning.
