Weakness of this AI

The unsupervised learning strategy implemented was able to adapt and solve the situation it was faced with: the Room Scenario Game. However through this experiment it became clear that there were some factors upon which that the AI performed weakly:

• It has difficulty with larger and larger amounts of data, as it becomes very difficult to map causes to effects.
  • To solve this problem a constant mapping process should occur in the background, identifying events with potential actions. Sensory data should be clustered together, rather than treated as discrete elements. It should also be possible to be quantitative with sensory data, to be able to compare how values compare to one another.
• The AI makes the same mistakes over and over again.
  • Making the same mistakes is not easily solved, as the software cannot disregard a failed solution permanently, as at a future date, the conditions may have changed and the incorrect solution becomes the right one. A possible solution would be to run 'tests' in the mind of the AI before performing anything, to check that the planned actions are still possible. Implementing a system where unsuccessful strategies time out is a simple, but flawed solution - it introduces more randomness, it doesn't solve the problem, and doesn't help the software learn.
• In this AI, there is no love! No obsession! No appreciation for aesthesia. Humans are fascinating because they appreciate ideas and actions based on very good things they associated them with in the past. An AI that would appear to be interesting should have these qualities.
  • Solution: the AI should favor:
    • Items encountered at a very young stage in the learning process
    • Items which it is attracted to (this should probably apply to complex things the machine gets along well with - for example, things that are not simple to learn about, but due to specific skills, or previous appreciations, they are easy to interface with). Attractive items would be ranked more highly in the decision making process.
    • Items that benefit the AI itself, over other things. I think the answer here is to teach the concept of material possession (to allow the AI to recognize assets as its own, where an owned asset would be an asset at the will of the AI), although Marxists would disagree. Possession is a jealous love, and in humans and animals, I feel, it derives from the need to find an exclusive sexual partner for reproduction. Taking this into account - is it possible to develop AI without making it reproductive? Possibly not. Reproduction seems a daunting problem if it is thought about in terms of the AI rewriting it's own code, however animals beings do not rewrite their own genetic code in reproduction, it is an automated process. Mating is not a process of sharing information (although it could be), and the resulting child should be nurtured by the AI parents, rather than absorbing information itself. To relate this back to the AI choosing things that benefit itself, the solution is to place a very basic demand that anything to do with the aim of creating AI children (who live long enough to reproduce) is slightly more important than everything else. As a consequence, all materialist intelligence should fall into place in the system. This basic desire would probably have to be part of the mapping cycle (mentioned above in the problem of large data), that constantly searches for a path from everyday items and actions to reproduction and survival.
    • The concept of exchange (the trading of goods and services for the benefit of all), however, should be a learned concept and not implemented directly.
• AI should exist in an active-state. It should constantly be thinking, and then able to choose when it wants to act. One of the fascinations of the Machine Room game was that it was so limited in how it should learn, and the human players had to adapt
  • This problem was demonstrated in the experiments, and concluded as: 'it appears that one of the most important aspects of good AI is its ability to choose wisely the timescale relative to its task'. This means that the AI must be able to execute tasks whenever it needs to (it must be aware of time).
• Humans did not develop alone - we emerged from communities, and so should AI. AI components should learn together.
  • If AI components exist in an active-state, they should be able to use their spare time to communicate with other AI's. This communication would be, at first, conducted in the same way they communicate with any other device.

Self Awareness

Is it possible to build an AI system that is self aware? It is possible that humans can only speculate on this question, because being self aware beings ourselves, we do not know what precisely that means. I theorize that human self awareness is part of a long chain of feeling and awareness that has existed since the beginning of the universe, and has been passed down through chemistry, to biology, and finally to conscious beings (humans) and mostly conscious beings (dogs, monkeys, etc.). I am not talking about God, but something called 'the spark of life'.
The aim of creating programmed AI is not to create self-aware beings, but to develop systems that appear self-aware but are constructed in an understandable way. These devices should then be able to perform tasks in the real world. This project appeared clever to the players of the Room Scenario game because they did not understand how it worked, and this demonstrated a perception of the players to its self-awareness.
Nonetheless, self awareness should be irrelevant to AI programmers, because it's not really part of the solution, but a possible outcome of software.
If the human brain has no software and it is simply a neural network of brain-nodes, then clearly, as nobody programmed it, and that it could possibly directly interface a human brain, then it might be self aware. This is a question for biologists.

This experiment revealed that it is relatively easy to create a basic system that once connected to a simple game, behaves in a productive and effective way. This reveals nothing about the potential successes possible once the software becomes more complex.
It is apparent that very slight changes in how the AI operates (the frequency of its interactions with the world, for example), changes how clever the system is perceived to be. Therefore, it is probably best to avoid anything that restricts the mind of the AI to keep it as open as possible.

Success of Unsupervised Learning

At the center of any general AI component, such as the Knowledge dll, there should exist unsupervised learning strategies, otherwise there is no hope for intelligence except for a repetition of man's skills. It was through unsupervised learning that humans became scientists, however, once these discoveries were made, they were taught to other people in the world. Therefore, AI should learn from a supervised perspective as well, through imitation and instruction.

Success of the Linear Memory

It is clear that the linear memory of the software could be represented in a tree structure, to speed up computation and to more logically calculate similar histories. This leads to an application that is primarily concerned with making decisions at points on a path, this is not necessarily the answer as it could be too structured. Perhaps we need a rough tree, a data structure with branches that don't always connect to each other and some that are not even part of the tree at all?

Entertainment Value and Practical Use in Computer Games

The Knowledge dll provided entertainment to players of the Room Scenario game, but was this entertainment a result of the AI's stupidity? For example, a human controlling the room scenario would probably be able to more quickly, come up with a strategy that would defeat the player of the game, simply because they are aware of how the room is arranged.

If the AI is made more intelligent does it loose its entertainment value? I believe the answer is negative due to two reasons:

1. The implementation of human values such as obsession and caring will make experience playing against this AI more interesting
2. As the AI becomes more clever, the games it is given should get more complex (especially games that do not rely on logic)

Good AI processing takes enormous computing power (the larger the memory, the greater the time to search - if processing video, this AI would take a very long time to accomplish anything).

The question: "given a computing device of infinite capability, is it easier to create AI?" leads to the statement:

"calculators can't be intelligent, therefore there must be a lower limit for the capacity of intelligence - perhaps the more simple a device, the more logical it has to be?"

Does this have anything to do with the separation between human beings and animals? Are there more points of separation?
It will be interesting to see how these questions are answered in the next 100 years.

To be more practical, the AI must be able to continually think on its own in its own thread on the computer operating system that it is running on. If the Knowledge dll implemented as part of this experiment worked in a computer game (the Room Scenario), it seems that the AI studied here is suited for computer games, although perhaps it's not a very efficient solution.

Final Remarks

It is true that the knowledge dll could probably learn to play tic-tac-toe, or even chess. However, it would be a terrible player because it cannot think properly for these situations, except basic logic. This is AI designed for real life, AI that if extended, and built into a computer game, could operate characters who you can really feel to be your friends. It is also aimed in the direction of household robotics. A household robot doesn't need to know how to perform mathematics, but it does need to know how to organize a house - a problem that is not terribly logical.