I know there are many debates about the definition of intelligence. There are many conjectures about what it is, where it comes from, and how to measure it. There are elusive references to hidden variables, such as the mysterious “G”. It seems to me that the ambiguity of defining intelligence is at least part of why we have not yet managed to create truly intelligent AI. How can we program something if we don't even know what it is we are trying to program? To me the difficulty of defining intelligence seems perplexing. Can't we boil this down to something concrete? Well here goes nothing; Here's my extremely concise definition of intelligence:

The capacity to adaptively refine behavior.

So before I get torn to shreds or mocked for my simple naivety, let me break this down. What is behavior? I consider behavior to be actions and outputs, such as twitching a muscle, activating a servo, displaying a character (or even a pixel) on a screen, sending a data packet, or just about anything an organism or system can actually do. What do I mean by refine? I mean to make more specific, more detailed, more particular to the given circumstances, conditions, or situations. And what do I mean by adaptive? Behaviors are adaptive if they serve the interests of the organism or system. In the case of an animal, behaviors are adaptive if they help it to survive and reproduce. For an AI, adaptive behaviors could be those that bring about the fulfillment of the AI's designed purpose. An example of an AI robot's purpose could be “To collect recyclable trash without disrupting the interests of other benevolent or benign agents”.

To me, this simple definition of intelligence seems obvious. It includes everything that we intuitively consider intelligent, and excludes that which we do not consider intelligent. A rock does not adaptively refine its behavior. A single cell bacterium does not refine its behaviors. Its behaviors are purely reflexive, fixed. Its behaviors may be well adapted, and through evolution its descendants may acquire refinements to their behaviors. But we do not consider an individual bacterium to have intelligence. It is not until an organism is capable of learning that we start to see the first glimmer of intelligence. A sea slug that learns to recognize stimulus patterns that indicate that it should retract its gill in anticipation of some potential harm has adaptively refined its behavior. It doesn't just retract its gill reflexively to any stimulus. It has learned to discriminate, to be more specific about when it retracts its gill. A cat that steps on a hot stove will never again step on that stove under any circumstances. The cat has some intelligence. If it were even more intelligent, it might refine that avoidance behavior to be even more specific, so that it only stays off the stove when the stove is turned on. The abilities to perceive, to learn to recognize patterns, to learn new sequences of behaviors, to focus attention, to predict outcomes, to plan, even to imagine, all of these functions serve the organism's or system's ability to adaptively refine its behaviors.

So what do you all think of my definition? What critical errors in reasoning have I made? What does it leave out? Is my definition so pithy and broad that it is useless? Will such a simple definition of intelligence lead an AI researcher down the wrong path?

~Dana Ream

Hi Dana,

I think that is a pretty good definition, but the questions I would have about your definition is:

• "who" has the behavior?
  
• "where" would they behave?
  
• "why" would the behavior be adaptively refined? What is the "purpose" of the behavior?
  

Maybe the answers to these questions are implicitly implied by your definition, but I think it would be better to make them more explicit.

The best definition of intelligence I have heard was from the 2011 Singularity Summit in San Francisco. Shane Legg gave the following definition of intelligence in words and also in a formula that is suitable for testing machine intelligences. The word definition is roughly:

• Intelligence measures an agent’s ability to achieve goals in a wide range of environments.
  

The formula definition from the summit is shown graphically here: http://yfrog.com/n5j3... and a paper showing the full theory of a measure of machine intelligence is here: http://www.vetta.org/... .

So, this definition automatically implies the following:

• There has to be an agent.
  
• The agent has to have goals.
  
• The agent has to be able to try to achieve these goals in a wide range of environments.
  
• There has to be multiple possible behaviors that the agent can choose between.
  
• There has to be a way to determine if the agent has achieved his goals.
  

By the way, this book: "The Genius Within..." by Dr. Frank T. Vertosick Jr. ( http://www.amazon.com... ) has convinced me that colonies of bacteria (not an individual bacterium), the immune system and even cancer cells have some real intelligence given the right definition of agent, behavior, goals and environments.

Frank

To answer your questions:

•"who" has the behavior?
The organism or system that is purported to have intelligence.

•"where" would they behave?
They would behave in their habitat, in the case of animals. For some artificial system, they would behave in the environments in which they were designed to be able to achieve their goals.

•"why" would the behavior be adaptively refined? What is the "purpose" of the behavior?
I consider a behavior adaptive if it serves the interests of the organism or system. "Interests" include goals. For animals, survival and reproduction are the goals/purpose. Any other sub-goals ultimately serve these greater interests/goals. For an artificial system, the primary goals/interests/purpose can be whatever the designer chooses. But unless the designer explicitly programs every behavior the system will ever be able to perform, we would want the system to be able to improve its effectiveness by adjusting its existing behaviors and hopefully even learn new behaviors. New behaviors are really just new combinations or uses of things the system could already do. These new combinations of actions or using old actions in new situations would be refinements in the system's behavior.

As an example of adaptively refining behavior, think about learning to type. Before you learn to type, you have to hunt and peck at the keys. But with time, you learn that it is quicker and easier to keep your hands in certain positions that put all the keys within reach of at least one of your fingers. With more time, you become more specific with which fingers you use for certain keys, and you become fluent enough that you no longer need to look at the keyboard. Now the act of pressing a key with your finger uses the same muscles that you have been using since you were an infant. As an infant, your use of these muscles was very clumsy (not yet refined). But the muscles worked, and you refined your ability to activate them in coordinated ways to grasp objects, scratch yourself, poke your friends, etc. When you learn to type, you further refine your use of these muscles into even more coordinated and specific combinations of individual muscle twitches.

I consider each muscle twitch to be a low level behavior, from which all higher levels of motor behaviors are constructed. Increasing the specificity of when you perform a low level behavior, thus reducing clumsiness, and learning new situations in which to perform specific behaviors are what I mean by "refine behavior".

It seems to me that the definition given by Shane Legg, "Intelligence measures an agent’s ability to achieve goals in a wide range of environments" is too inclusive. It would include any animal that manages to survive and reproduce, even a single bacterium.

•There has to be an agent.
The bacterium.

•The agent has to have goals.
To survive and divide.

•The agent has to be able to try to achieve these goals in a wide range of environments.
There are some pretty robust bacteria out there that can survive in a wide range of environments.

•There has to be multiple possible behaviors that the agent can choose between.
Divide or go into stasis or move in the direction of some chemical gradient.

•There has to be a way to determine if the agent has achieved his goals.
Did the bacterium produce many descendants?


But a bacterium's behaviors, its "choices", are reflexive and always done the same way in any given situation. It may even be able to take into account complex combinations of stimuli. But it cannot learn that there are exceptions to its decision making rules. If it could add a new exception to its decision making rules, that would constitute a refinement. If that new exception helped the bacterium to better achieve its goals of survival and reproduction, then that refinement in behavior could be considered adaptive. Only then would I be tempted to call the bacterium (or anything else) intelligent.

~Dana Ream

I don't agree with this 'simple' definition:
"Intelligence measures an agent’s ability to achieve goals in a wide range of environments."

If an agent is adapted to achieve goals in a wide range of environments,
that agent may be less adapted to achieve a particular goal in a particular
environment, compared to an agent that is well adapted for the specific goal
in the specific environment. There is always a trade-off.

Intelligence is domain specific. Sometimes we want specific intelligence, and
sometimes we want general intelligence, taking a hit in some specific intelligences.

My simple definition of intelligence is, "If you agree with me, you are intelligent,"
that is, "Intelligence is, simply, that which an intelligence test measures."

I'll expound more later.

"There is always a trade-off." This may not be correct. It is claimed that children who learn music are better at learning other subjects, than children who don't learn music. When is learning synergistic, and when does it interfere, is something I have to think and read about more.

Dana:

On your example of the bacteria, if the goal of the bacteria is to survive and reproduce, then you have to let it reproduce multiple times over an extended period of time to measure it's success at reproducing and if you do that, you end up with a bacterial colony that can use evolution to modify it's behavior and I would indeed define an evolving bacterial colony as a kind of intelligent agent. See the book The Genius Within...

Al:

The purpose of the definition I quoted was to allow a comparison of the intelligence of agents. To do that comparison you must define the domain of environments. If you have either a wide range or a narrow range of environments you can still compare to agents by which one is more successful. But it is inappropriate to compare one agent on a narrow range to another agent on a wide range of environments...

I agree that bacteria have intelligence. Intelligence seems to be a general property of the universe. To drop a name, Dr. Michio Kako once said that even a thermostat has a 'degree' of intelligence. Is that intelligence comparable to human intelligence? Yes, if you define the test: the intelligence one is talking about needs to be tagged with a particular test, otherwise intelligence is not comparable.

Let's define the test goal as "survive an asteroid strike." Depending on the time frame and the size of the asteroid, as parameters of the test, bacteria could be more intelligent than humans, or humans could be more intelligent than bacteria, or they are equally intelligent or unintelligent. Let's hope we are more intelligent and figure out a way to survive on other planets before a 'planet killer' strikes ;)

As a digression, here are some interesting articles about bacterial communication and energy transfer.
http://www.newscienti...
http://www.newscienti...
http://www.newscienti...
http://www.newscienti...

Then again, it has been suggested that life on this planet originated from bacteria from another planet. Now, who are the true rulers of the universe?

This is my simple idea about artificial intelligence:

AI is a computerized system that is designed to achieve sets of goals whether it is domain specific or in multiple domains. It has program algorithms that work toward the desired goals or actions. It has data resource specific to the domain or domains that the system can use to generate outputs or actions. Moreover it has ability to collect new data, able to verify the data quality. To be more adaptive to the environment, the system needs to be able to form or organize new concepts or knowledge from changing environment; and it needs to be able generate new options for outputs in order to be adaptive, able to learn the mistakes and failures as well as successes of the actions the system generates - able to evaluate the results. The system needs to be able to accumulate new knowledge and achieve the desired goals in new situations.

To answer Dana's comment about bacteria's intelligence:
Bacteria can mutate into a new strain that resists to antibiotics that kill them. I consider this a kind of intelligence since they can adapt and survive in hostile environment.

Nalini

Intelligence is a fundamental (reductionist) property as well as an emergent (holistic) property of matter.

Example: The intelligence "Intelligence[ cut a piece of cloth ]," where "cut a piece of cloth" is the test tag, is not a property of scissors ALONE, nor is it a property of human beings ALONE, however, it emerges as a property of the combined intelligences of the "scissors + human" system.

Yes, I am saying that scissors have a built-in intelligence that humans don't have, even though that intelligence was enabled by a human!. (No wonder that the concept of intelligence is so simple, yet needs international conferences of scholars to understand.)

Now, what is human intelligence? It is an emergent property of the bacterial intelligences of the "bacteria" that human beings are composed of.

As an aside, I am debating whether one needs to tag intelligence as "Intelligence[ D, t, T ]," where D is the domain, t is the training set, and T is the test set, and where each is a set (ordered set in the case of t and T) of multi-dimensional (or even Hilbert?) vectors, each consisting of a stimulus part and a response part. What do you think?