When designing and implementing autonomous agents, it is important to keep in mind that recognition is not the same as understanding.

Recognition

The definition of recognition is “identification of someone or something from previous encounters or knowledge.” The etomology of the word is from the Latin “to know again”. Recognizing things like commands in programs or hyperlinks in web documents is essential for an agent to accomplish an action.

Importantly, agents can recognize elements in their surroundings (real or virtual) only when they are seeing it again. Agents of all types, including humans, must know something about a thing ahead of time in order to recognize it later. This has important implications for how you can program (or train) an agent.

Conversly, not knowing a thing ahead of time means it is likely agents will not see it in the future. We’ve probably all had this experience ourselves. Without knowing what to look for, it is difficult to recognize objects around us.

Recognition is required before understanding.

Understanding

The defintion for the word understand is “perceive the intended meaning of (words, a language, or a speaker).” Understanding something implies a grasp of the meaning behind it. For humans, understanding the meaning of a word or phrase is essential for communicating. If you don’t understand the words I am using, you are unlikely to grasp the meaning of what I am saying (or writing).

Meaning is tied to intellect and reasoning. Our understanding depends on our intelligence and our ability to reason about what we see and hear around us.

We sometimes talk about “intelligent agents” or training agents to “understand” new commands. But, in reality machines do not grasp the meaning of things. They cannot reason about their surroundings. This level of agency is usually called “artificial general intelligence” or AGI. And, as of this writing anyway, AGI is still only a hypothetical possibility.

The good news is that, even without understanding we can enhance machine agents by way of inference. But I’ll save that for a future post.

Leveraging Recognition

The kind of experiments I’m working on with the AMEE Project are the ones that leverage an agent’s ability to recognize elements in the environment and, when appropriate, act on those elements. That means designing and implementing agents that have the ability to recognize things. That is, the ability to “see again” things it already knows about.

That means I’m desgning and building more than just agents, I am also creating vocabularies (sometimes called ontologies). These vocabularies act as the engine of the agent. They provide the a priori information agents need in order to be successful in an environment – in order to recognize trhings.

And that’s an important step. Creating environments that can contain recognizable elements and agents can that both recognize and act on those elements is what the AMEE Project is about. However, recognition is only part of the story. Questions like “How does an agent recognize something?” and “How does an agent know what to do once it recognizes things?” are all key to moving beyond static programs and toward programs that can act on their own – that operate as autonomous agents.

– @mamund