Logic, emotions and AI – Are emotions still thought of as unpredictable outbursts of uncontrollable behaviors? Or might they have some value, as aids to understanding human expression, kind of like icing on a cake? They contribute but are not absolutely necessary. I see them in 2 forms. 1- Expressed as feelings, but an expression of concerns regarding homeostasis. 2 – A system of alarms. A means to communicate without the knowledge of language.

It seems to me that logic or rational thought serves to maintain an entity’s physiology within the limits of comfortable homeostasis or without tripping the alarm system. It also seems to me that the more intelligent an entity is the more resources it would have to avoid dangerous circumstances. To me these are questions about survival that provide motivation regarding the best way to accomplish survival tasks.

It has been suggested that other methods beside close examination of the makeup of animal internal psychological evolution could achieve equal intellectual results. I am very curious regarding how and why this entity will accomplish this intellect without the benefit of self-interest.

Lex,

What I would do if I were trying to architect an approach to synthetic intelligence is to look to rhetoric free methods. In an earlier post, I used that term in jest, but it now seems to hold some deep meaning. We don't know enough about the human mind to make much more than guesses as to the most general shape of most of its gross functioning and mechanism. Things like "reason" and "emotion" are somewhat arbitrarily bounded regions of overlapping and interdependent systems that may in fact be only attributes of some larger, or aggregates of many smaller systems. We just don't know enough about the brain. Certainly not enough to begin the necessarily detailed process of the actual construction of a brain analog. Same goes for any of the brain's supposed parts or systems.

So, if the intrepid researcher, hacker, nutcase (myself included) is still interested in this Atlas-like pursuit, that person had better find a starting point that allows the possibility of success without detailed knowledge of a parts list or for that matter the arrangement the parts need to take. Again, we don't know enough to attempt this task from the perspective of an architect or builder.

What this means is that we are forced to build, not an intelligence system, but an intelligence building system. And we are of course forced to do this from the acknowledgment that we don't know enough to build from knowledge. We must build a machine or system that can grow intelligence from not intelligence.

Luckily, we have a model for such blind construction projects, we call it evolution.

A properly designed evolution program can treat every part of the proposed construction goal as a black box or black box of black boxes. Remember the order of events here. Nature evolved the brain first, and then the brain tried to fit labels on categories of its behavior (reason, emotion, consciousness, sentience, etc.).

When Monica uses the term "model free methods" she is I am sure, admitting an acknowledgment of the fact that most of the top-down assumptions that preoccupy AI research are based on overwhelming ignorance of the very thing we are trying to build. She uses that term to put everything about intelligence into the category of "we don't know enough to build from understanding". In fact, she goes this admission one further and claims that these categories and parts of intelligence are arbitrary and their arbitrary-ness is and always will be an artifact of understanding and that understanding is an after-the-fact abstraction or result of processes that operate below anything that could be called understanding… the linear, static, serial tailings of many processes, none of which are in fact intelligent, emotional, reasonable, sentient, conscious, etc.

So… though it is most certainly true that the attributes and behaviors of the brain that result in what we experience as emotions have profound causal effect on what we experience as intelligence, it is also true that we don't know enough about emotions to design a mechanism that would either produce them or react to them. As such, we AI researchers must treat emotions the same way we treat every other attribute of intelligence… as a black box. And we must build towards a system that can both precipitate the emergence of systems (as black boxes) and precipitate the emergence of aggregation schemes that handle the complex interaction of these black box systems.

Getting caught up on, or attached to, any one of the many arbitrarily bounded and little understood black boxes that appear to result in human intelligence is a sure way to distract us from the restrictions ignorance places on the parameters of the project to build a synthetic intelligence. Our only choice is to admit ignorance and plan an attack on the problem that allows for success despite ignorance. Evolution.

Randall Reetz

Randall, Thanks for your reply. This is a journey and as such we must consider what directions to take. It seems to me that you have decided to wait for someone to construct an exact miniature model of the journey before you start to proceed. It is true that what I believe I am offering is more akin to a road map, but the question remains - Can we get there using the instructions from these methods?

To me, it seems crystal clear that needs, feelings and what is termed as emotions are at the heart of what motivates the animal world. Furthermore, logic and reason do not exist without motivation. The point I’m attempting to make is if we want to design intelligent machines and we don’t want to recognize the involvement of animal motivation ( too great a paradigm shift) or can’t (because of complexity) what are we going to do? Give up before we’ve tried.

It’s obvious that there is confusion regarding the necessity and the complexity of representing animal emotions within a machine. It can be accomplished by a slight shift in how we handle input data. Let’s say we use a thermister to sense temperature as part of a robotic system. Maybe it’s sensing the temperature of the room or one of its motors. Its resistance decreases as its temperature increases. Therefore the current through it increases. This is digitized and stored as part of some experience. Normal reductionest programming technique uses this kind of data for direct control of some environment. However, this could also be used to describe the beginnings of a system of how it feels regarding the entity’s thermal condition. This system would consist of many thermisters and tell a complex story of its thermal homeostasis. This kind of system among others is what the animal world uses to establish the meaning of wellbeing. It is also the information used in the formation of concepts. As concepts are formed the degree of wellbeing is therefore integral to and associated with the new concept. Because of or lack of this wellbeing association, an entity can decide weather it needs to make an adjustment. As more complex concepts are formed, the same test for wellbeing is performed. Logic and rationality come from the consistency of the performances of our animal input systems and the similarity of environmental circumstance that we share with other members of our population. As evolution has added layers of hardware capable of successfully handling greater complexities, the systems which provide the reason behind performing these complexities remains, as far as I can tell, the same.

Lex,

I am not saying that emotions are complex. I famously contend that emotions are as simple as a set of nested if/then conditionals – if hungry, eat… if girl walks in, stop eating… if lion walks in, stop kissing girl… if neighbor kills lion, give him the girl, etc. What makes emotions seem complex is the way their simplicity sets up bizarre (thanks Monica) cascades of memory association triggers in a huge brain.

Of course my "emotions as nested conditionals" theory is hardly as easy to build as it is to declare. I am convinced that intelligence (of the sort that humans would consider – remember that we seem to don't often recognize sentience in other animals with with complex brains) is probably dependent on machines close to the complexity of our own brain (100 billion nodes each with an average of 10 thousand connections). We have trouble building commercial jets with 100 million parts arranged statically.

I am not saying that emotions are impossibly complex. I am not saying that any one subsystem of the brain is beyond the complexity handling capacity of our engineering. But I am saying that the discreet design of a system this complex and this interrelated and inter-conditional and interdependent is probably best attempted as an emergent system. Instead of asking "How can I build a Redwood tree?" we might better ask, "How can I build a growth algorithm as meta-seed that can be tweaked through iteration loops to grow all manner of plant life?"

There is probably some threshold of the ratio of relative complexity between mind and engineering target beyond which the mind will not succeed. The bigger the project, the bigger the mind has to be. But this rule would only apply to the process of engineering, where the entire project needs to be conceptualized and manipulated comfortably within the brain before it can be built. But that is only if one needs to be able to pre-compute the entire arch of the project history. The blind evolutionary process needs no such capacity a priori.

The other advantage of the iterative evolutionary process is that it tends to build a great many functions from a small set of generalizable atomic structures. The resulting causal hierarchy of stacked grammars allows a profound elegance of symmetries that affords economic solutions to a wide range of problems and instabilities… the solution space is regular and transitive.

Not only does evolution work for the evaluation of tree growing algorithms but it works for the more basic seed designing process as well. And you don't have to start blind. You can introduce all prior knowledge and theory as environmental constraints that will then contribute to and modify the seed's own internal fitness metrics. Evolution has a way of internalizing as abstraction pattern, the environment that surrounds the entity evolving. Come to think of it, the mind "learns" by the same mechanism. Building a more and more salient abstraction of the environment around it.

Evolution is the very opposite of the homunculus ("construct an exact miniature model of the journey before you start to proceed") you mistakenly ascribe to my approach. I have a lot of philosophical opposition to top down engineering projects, especially when the target project approaches the complexity of the entity doing the engineering.

Randall Reetz

I must admit that I have not read and understood all the multiple replies in detail on this topic, but I thought I would just add my $0.02 on the topic of emotions as seen by my theory of the Primary and Symbolic Consciousness that exist in every human brain. The Primary consciousness is also shared with all higher animals and emotions are part of the Primary consciousness system. Both Primary and Symbolic consciousness have built in goals, such as procreation of children, staying alive, etc etc and each of these goals get broken down into lower level goals finding a willing individual of the opposite sex with whom we can procreate, finding food and water to stay alive, etc etc.

In this view, emotions are the positive and negative feedback mechanism the Primary consciousness uses as it attempts to achieve it's goals. If the goals are being met, emotions of happiness and joy are experienced to induce the being to do more of that again. When we are blocked somehow at getting our goals met, we use anger or frustration to help us remove the blocks to achieving our goals. So, in this view emotions are tightly tied up with the goal meeting mechanism of the primary consciousness.

Now in humans with an additional symbolic consciousness, the emotions from the primary consciousness are experienced as being in the body and they serve as a communication mechanism between the primary consciousness to the symbolic consciousness. This is how the primary consciousness tries to influence the symbolic consciousness to help it meet the goals of the primary consciousness.

In humans, there are even more emotions than in animals because the symbolic consciousness has the ability to continually "talk" to the primary consciousness and to try to impose it's many goals on the primary consciousness which then results in more emotions as the goals are either met or not met. These emotions are thus the result of the various obsessions and other goals of the symbolic consciousness. For example, the symbolic consciousness might continually obsess about a member of the opposite sex even though they have already turned us down. The primary consciousness would give up and go on to find another candidate but the symbolic consciousness just keeps obsessing and obsessing after the one that got away. There are just a lot more emotions due to the symbolic consciousness since the symbolic consciousness does not know how to accept the things it cannot change, it does not know how to change the things it can change and it does not have the wisdom to know the difference.

That's my $0.02.
Frank Heile

Frank,

That is interesting. But I wouldn't need to build an emotion machine into a synthetic intelligence because I would assume that emotions are emergent properties of any intelligent system. It just seems so silly to build a happiness routine, or a sadness algorithm, or, or, or. Reminds me of the animatronic ducks that were built in the 1600's in france and austria. Intelligence deserves, demands, more than artifice.

Randall Reetz

Frank, Thanks for joining the discussion. I think your spot on regarding your theory. At this early stage of this discussion I was trying to focus on two aspects of emotion and logic: 1- their physical connection with the body and 2 – their relationship with each other.

In the past AI researchers noticed that the more complex or intelligent ideas became the more they seemed to be logical. Intelligent people did not through tantrums to make a point. They conducted themselves in a rational, logical manor. The conclusion was that if a machine were going to be intelligent than we were going to be dealing with logic. This conclusion was further supported be research regarding brain maps. I’m sure your aware that there are areas of the brain which deal solely with emotions. This has further separated logic and emotion and served to cloud the issue. Emotions are viewed as simply adding spice to animal functionality and, in the extreme form, as dangerous and as you mentioned primal. This is not to say emotions aren’t these things, they are. But it seems to me that this perspective is incomplete and a lot so. What I get from your discussion is a perspective formed at a much higher level of conceptualization than the early stages I am referring to.

However, I would be interested in what you feel are the mechanisms which separate the Primary consciousness from the Symbolic Consciousness. I have felt that early stages of development were obviously nonverbal. At this stage we are learning about all of the dimensional aspects of our environment. We are learning the meaning of the physical world and we are forming a non-verbal thought process. This is why we have such a problem describing what we do when we think. Anyway, Symbolic Consciousness seems to coincide with language.

Also, I am interested in the communication between the Primary Consciousness and Symbolic Consciousness, I see this as an ongoing com event as well. I would add that within the Primary consciousness mechanisms are constantly assembling concepts while others are monitoring for any emergent needs. The Symbolic Consciousness appears to be almost simplistic in comparison. It decides whether or not it is better to do, eat, etc this or that. It seems to direct the Primary Consciousness to present the concepts which will support any homeostatic needs it deems necessary. If it has time, that is to say if homeostasis is normal, than it can start a creative process or a thinking process of what might be important in the future. It is simply a decision-maker that uses words, but, this is a topic of another thread.


Randall, What if it felt sad if it weren’t serving you up to your expectations. Or if it weren’t pleasing to the device in some way, why would it serve anyone at all? The real question is how are going to get these things to do what we want? I think the answer is found in the question why do animals do what they do. And there are answers. Feeling sad is just a low-level alarm. Regarding a machine it may or may not need sadness. The matter will be totally in our hands.

Randall, I welcome your critical judgment. You do remarkably well with it. Sincerely, thank you. As I have mentioned before, if my path doesn’t lead where I expect than there is no need to go there. If you should see any of my arguments as some form of wishful thinking, don’t hesitate to point that out. I am about moving forward. However, I have not painted the picture in its entirety. This topic has only served to organize what I see as necessary to begin.

The complexity of the devices you refer to above, I agree are way to complex to be accomplished in the way you describe them. That is why I suggest we don’t do it that way. The reason I believe the algorithmic approach you describe would be so complex, is that at the starting point the conceptual level is complex. I see a much simpler starting point than the one you describe. Your suggestion to ““Instead of asking "How can I build a Redwood tree?" we might better ask, "How can I build a growth algorithm as meta-seed that can be tweaked through iteration loops to grow all manner of plant life?"” is, pretty much, what I am suggesting to do. The meta-seed you refer to is encapsulated in what BF Skinner said regarding all animal behavior. “All successful behavior is tended to be repeated.”. They can train chickens to play the piano using this. They can train sea slugs to avoid shocks. This is at the heart of motivation, but, the question AI needs to ask is: “What does successful mean.” What it means to me is to repeat whatever action is necessary to fulfill whatever out of wak need regarding homeostasis. Granted simply doing this does not produce an intelligent device. No where near. This assumes an ability to form concepts and a very simple control modual. As its concept handling ability is increased its intellectual ability will also increase. The degree of intellectual ability will be a contest between its ability to form concepts in a timely manner and the interval between needs fulfillment. Granted this will take needs structure tweaking. The complexity of the conceptualization modules concerns me. I have some ideas but if you want to poke holes in this model this is the place. I think a lot will depend on the kinds of input hardware that is selected.

Ok Lex,

So we agree on the need to evolve systems as complex as those capable of intelligence. Why then do you see such a need to be an apologist for, a cosmic gardener of, emotions? Its as if you are saying, evolution is fine but it won't result in the mechanisms required for the emergence of emotions… requiring instead, some top-down tinkering. I don't get it. What is it about emotions that you feel won't arise naturally from any intelligent system? I smell rhetoric.

Randall Reetz

“So we agree on the need to evolve systems as complex as those capable of intelligence. Why then do you see such a need to be an apologist for, a cosmic gardener of, and emotions?”

Randall, I think you are mistaking my request for critical thinking for an apology. I do feel that I am walking a fine line regarding the perception of this subject. The subjective nature of emotions makes a discussion of them difficult within a scientific framework. Most of science questions their relevance and fails to see what they provide. This is especially true of AI, where logic is king.

“Its as if you are saying, evolution is fine but it won't result in the mechanisms required for the emergence of emotions… requiring instead, some top-down tinkering.”

I don’t see your thought progression here. If I understand you right, I think that I share your view that evolution extends itself into all aspects of the universe. Human or otherwise. You have said that all things evolve and I agree. Its presents is felt in all things. My point regarding emotion does not try to avoid this and if it appears that I am deviating from the path of evolution than that is unintentional. To me this is all about getting a chunk of metal to be self-actuated. I don’t care if it takes emotion or emulsion; it takes what it takes. Earlier, you stated that you didn’t have a need for a machine to react with sadness. But isn’t sadness a statement of internal condition, albeit vague. It functions as a specific kind of alarm. As I have said we will be in control of the kinds of responses these machines will be capable of producing. Our judgment will serve to control the evolution of this system. At this time I don’t believe hardware will be capable of duplicating the full spectrum of human intelligence; however, I think it will be interesting to see how far we can get.

“What is it about emotions that you feel won't arise naturally from any intelligent system?”

I feel emotion does/has arisen out of nature through natural selection and it is a requirement of motivation and therefore a requirement for any intelligent system. Intelligence allows us to control/adjust our environment to better serve our needs. As you have pointed out, it allows those with intelligence to be much more competitive for the limited recourses available to them. Emotion is a symptom of our needs. And key here is that there would be no intelligence if it did not serve those same needs. To understand emotion is to understand how to foster intelligence.