AI agent tutorial - agent classification | AI Agent types

Key Takeaways

This video is all about the different agent types or agent classification. We're going to start with the different components of an AI agent. Then we're going to look into the different features of an AI agent. Then we're going to move on to the core classification itself. And that can be broadly classified into five types based on their level of intelligence and their interaction with the environment. You know how autonomously they can interact with the environment and make their decisions. So let's get started. So let's start by looking into the different components of AI agents. The AI agent has an environment where it interacts with. So in case of LLM, it will probably be the user who interacts via an user interface. But there will be some form of environment. For example, in case of an autonomous vehicle, the environment is pretty much the people walking around the cars and everything that's there in the road. At the heart of the agent, we have the model which is bringing in all the intelligence. But at a basic level, we need not have a model. It could be just a condition action loop as we'll see in in the simple form of agents. But the agent definitely has a persona and it's self-directed. It kind of makes some sort of decisions. It direct itself to do something and then it has access to some tools. For example, in case of LLM again, it can have access to browser. It can have access to calculator to do some some computations and it also has some planning element meaning that you know it can come up with few actions action Action B and then decide okay action B is better than action A and it decides to do action B instead of A and it also has memory to store the intermediate states as we will see in a different type of system there is me memory that's needed in order to make the agent more sophisticated talking of memory can be different types of memory you know consensual memory episodic short-term long-term memory. So these are the main components if you like of an agentic system. These are the sort of essential components that make it more sophisticated. And talking of the key features of the AI agent, the AI agent will have a certain level of autonomy and the more autonomous the more sophisticated the AI agent will be and it'll also be goal- driven meaning that it will work towards attaining some goal. It'll have some goal where it needs to reach or you know what it needs to achieve and it also perceives from the environment. It is adaptable meaning that you know an advanced AI agent can adapt itself. It can learn it can change itself and it can kind of evolve into the better agent if you like and it can also be collaborative. So advanced agentic system like you know multi- aent systems they can collaborate like you can have more than one agent just simply passing message between each other and achieving a task that is quite challenging. So those are the kind of key components of AI agent. So let's look into how we can sort of classify the AI agents. One way to look into classifying them is based on the level of intelligence or the level of interaction that they do with the environment. And we can also consider the level of collaboration that goes between the agents when you have multiple agents in the system. So broadly we can say that they can be classified into simple reflex agents, model based reflex agents. As it gets more sophisticated we can have goal-based agents, utility based agents and we can also have learning based agents which can be highly advanced sophisticated agents. So let's look into each of them one by one. So let's start by looking into a simple reflex agent. The agent of course has uh environment that it interacts with. The agent receives the signals or the percepts from the environment. It could be you know some data that is perceived by some sensor and it goes into a decision-m system because this is a simple reflex agent. The decision-m system is not very sophisticated. It may not even have a model. It could just be if else condition. And based on the condition and based on the sense data that we make a decision as to whether we need to act or whether we do not have to act. So if it decides to act then the action actually impacts the environment in turn. So think of a thermostat where you have a air conditioned room and you are constantly perceiving the temperature of the room using thermometer. So the measurement of the temperature is what you get from the sensor and you just have a condition here saying that you know if the temperature is below 22° or 23° then actually turn on the air condition and if it's not the case then don't do anything. So this is a simple reflex agent and so a simple reflex agent is applicable in predictable environments where you very well know what exactly will be there in an environment. Then you can deploy a simple reflex agent and the rules can be very well defined with some logic. Again think of thermostat where you can actually define a temperature and say that you know whether we need to cool or not and good examples of simple reflex agents or you know traffic light systems or your automatic doors where you know if there's a person before the door then open if there's no person just close the door. So all these come under simple reflex agents. These are very basic agents. So let's move on to something slightly more sophisticated which is modelbased reflex agents. So the main change here is the introduction of memory otherwise called the state in robotics. So we have states which stores the sensor data the action data of whatever has happened in the past and on top of having conditional action rules. You may also have a a simple machine learning system that feeds into the decision- making. For example, the ML system can um can predict something based on the sensor data and say to proceed with the action or not to proceed with the action. But in any case, we do have a state that is religiously tracked when it comes to modelbased reflex agents. So the model based reflex agents, they act in slightly less predictable environments. So a good example is the robotic vacuum cleaner that we have in our houses these days. So the vacuum cleaner should store the areas that it actually covered in the room so that it need not have to revisit those areas. Though it is a predictable environment, it may come across some you know obstacles when it's moving around that it did not perceive or did not map when it was actually put in the room. So these are the kind of examples that are actually model based reflex agents. So the next level of sophistication with agents is the goal-based agents where we introduce goals. So these agents are quite intelligent and they work towards achieving a goal rather than simply working with some state. They also have state but they now have goals. So whenever they perceive the environment, they come up with a plan first and using the plan and using a highly capable AI model they actually try to execute some action or in fact they come up with a few action items they execute the action that is the best and once they execute the action the loop continues till the goal is achieved. So in case of goal based agents adaptation is quite essential because you know you have a goal you are working towards that goal and you first come up with a plan you're doing you're coming up with action Action B etc and you're weighing as to which action to perform and you execute it but then the environment gives you you know a curveball and you decide that that that action is not great then you have to kind of redo the plan. A very good example is that of autonomous driving. You know you give a destination to the car where it needs to go and then the the vehicle starts driving but then it can come across road closures. It can come across you know obstacles on the road. Then it has to change the plan. It has to even you know do some diversion. It has to completely come up with a new plan and then execute a new plan of action. So that that is a very good example of a goal-based agent. you can go ahead and build even more sophisticated agents and those are utility based agents. So the utility based agents have the powerful tool to evaluate. So basically utility based agents they don't just achieve the goal but they actually use the best possible way to achieve the goal. Basically they look for optimizing based on your criteria and that criteria is defined by using a utility function. So you provide a utility function to the agent and you define the utility function and you define a way to somehow quantify to say that this goal or this way of achieving is better than that way of achieving your goal. The agent decides between action A and action B and thinks that action A is better because it has a rational to decide which is the the quantity or which is the evaluation it does and so goes for action A over action B. So that is what I mean by multiriteria decision- making. So it has a multiple criteria to decide which is the best thing to go for and there's also trade-off between competing goals. You can have multiple goals. A very good example is that of autonomous driving. So if you go in Google Maps and put a destination, you can optimize for fuel efficiency rather than you know just getting to the destination and you compromise over speed. So that is a good example. And if an autonomous driving car decides it wants to optimize for fuel efficiency, that car is utility based agent rather than a goal-based agent. And another good example is that of financial portfolio agent. And the trade-off is there could be multiple ways to get rewards or returns. But the agent has to evaluate each of the available options and then decide which one is the best balance between risk and return. So that's another example of utility based agent. The last but the most sophisticated uh AI agent is that of learning agents. So in case of learning agents, you have critic. So basically whatever the action that the agent does, it is sensed back from the environment by using a sensor and that's fed back into a critic and the critic in turn feeds a learning element and the learning element is tied to a performance element. Although the learning element can evaluate different criticisms that it gets and kind of improves because it can evaluate the performance based on different parameters that it changes and finally in order to do that it generates a new problem. It it evaluates it and it goes through this loop until the entire agent itself evolves into a better one and then finally it does the action and the action in turn is executed in the environment and the loop goes on and on. So a very good example of learning agent is the autonomous robots itself. So they become adapt at navigation. For example, the terrain in which these autonomous robots move is going to change constantly because you don't know what is going to come as obstacle. You don't know how the terrain is going to be, you know, a sandy terrain or it's going to be a slippery terrain or whatever be it. But the robot has to kind of adapt itself. And it also needs to interact with humans and and humans are quite varied and the inputs that they get from humans can also vary a lot. And the next example is that of recommended systems. Again, users or humans here. So you don't know what sort of inputs you're going to get into the recommended system, but the system needs to become better at recommending by learning from the user's behavior. And each user has a separate behavior. And so the system needs to adapt per user. So these are a couple of good examples of learning based agents. So to summarize, we saw about these kinds of agents. Simple reflex agents are fully observable and good examples are automatic doors or traffic lights. Model based reflex agents are partially observable are good in partially observable environments and a good example is robot vacuum cleaner. Goal based ones are involved when you want to come up with a strategy for kind of achieving a defined goal. A good example is autonomous navigation of a robot or a autonomous car. Utility based agents are needed when you have a specific criteria or you want to optimize for decision-m. A good example is that of financial portfolio management agent. And the last one is that of learning agent which is needed when your environment is evolving and it's quite dynamic then you need recommended systems. So when your environment is evolving and it's quite dynamic then you need a learning agents that can evolve along with the change in environment. Good examples of recommended systems. So that pretty much wraps up our review of different agents that are available for us in the literature. Let me know if you want any hands-on on any of these. Leave a comment below if you want to see any hands-on videos and I will endeavor to get some hands-on videos as well. Until then, I'm signing off and I will see you in my next one. Take.