The current M-Logic Machine implementation is somewhat more complex than the example given above. In it, frames use around ten channels to structure data. The searched patterns are four frames long, instead of just two. A dozen of predictive heuristics are used, along with half-a-dozen recording criteria. The heuristics are subjected to a dominance mechanism similar to the one used for scenes in the Dominance-List Memory. This dominance memory is limited to less than two hundred scenes, and is typically too small to record all the possible sequences of frames. An off-line (sleep-mode) abstraction mechanism helps locate the relevant channels. Anyway, having seen in some detail how the MLM works in a way that implements reinforcement learning, we can better understand some aspects of M-Logic foundations.
1. Pleasure and Pain:
We saw that the M-Logic Machine uses heuristics that repeat a past motor action when the a cover-set value tagged "pl" is found in the next frame. It's very important to notice that the MLM does not repeat the action because "pl" means "pleasant". It's the other way round: we call "pleasant" the cover-set output that was chosen by the heuristic to trigger a repetition of a past motor action. On the other hand, "pain" is any cover-set output that is used by a heuristic to trigger evasive motor actions. The heuristics just use specific channels of the frame (or scene) structure, the "driving channels". The measurement cover-set instrument that outputs pl or pn may do so based on vital aspects of the agent, like the increase or decrease of the internal supply of energy or the structural integrity. If the link between these outputs and the heuristic is inverted, the machine will simply cease to exist in a hostile world. This selective process implements evolutionary learning. In principle, all kinds of measurements can be mapped into the "driving channels". People with synaesthesia may elicit or reject some types of food because they like or not the colours associated with the tastes. This can generate a nutritional lack of balance. The "right" mapping is therefore not based on any particular meaning (or descriptive tag) of the cover-set output (this would import into the MLM the symbol grounding problem). It is rather a natural consequence of a survival constraint. The machine does not see tags. They are used for descriptive purposes only. The nature of the links is purely spatial. This does not exclude the possibility of combining spatial features to further restrict the possible links (for instance a key that matches a lock).
2. Truth and Falsehood:
The same idea applies to the notion of truth and falsehood. Truth is the configuration of the working memory that is used to pull up memories in the dominance-list memory. If this relationship is altered under the selective pressure of a hostile world, the machine perishes. Hence the aphorism "truth is life".
Of course, in a setting where no survival constraints are present, it's totally irrelevant what is chosen to be repeated or avoided, pulled up or pushed down. The notions of pleasure, pain, truth, falsehood, good, bad, etc. become irrelevant and undistinguishable. Giving meaning to these ideas requires the M-Logic Machine to be inserted in a hostile world. A world is hostile when it eventually eliminates the agent that interacts randomly wih its environment. Since the agent's inner processes may nevertheless persist if they are replicated fast enough, there are also supra-individual dimensions to the notion of world hostility (and intelligence).
3. Goals
Goals emerge from the heuristics used and the definition of what is pain and pleasure. These definitions depend on the sensory modes used, and the sensory modes used are often triggered by the survival needs of the machine. For instance, if we feel hungry, pleasure is eating, and cinematic memories for eating will be recalled. The knowledge of what came just before provides theories that may link the agent's present situation to a future satisfaction of its needs.
Learning not only helps the agent to adapt to its environment. The obvious impact of humankind on Earth's ecology illustrates the principle that all learning also strongly contributes to environmental change. This is another reason why the M-Logic Machine does not seek optimal lines of action, in favor of flexible and fast adaptations to ever-changing conditions.
4. Extensive Mapping
STM ans LTM cinematic memories have a small number of frame channels, much smaller than the number of possible classifier intruments that provide recordable information. The MLM needs a set of orienting reflexes to trigger the adequate sensory modes for the different situations. This is another aspect that is subject to evolutionary learning. The inference mechanism is unaware of the nature of the incoming sensory data, since all transmission lines and pulses are similar. To achieve domain-specific learning, distinct sensory modes must be recorded in distinct memory areas. This implies an extensive mapping of sensory modes into different areas of the LTM. It's the mapping that gives meaning to otherwise undistiguishable transmission lines and signals. Because of the non-numeric nature of this data, no means or medians can be calculated, and no expected utility can be evaluated. The dominant data is ultimately closer to the mode of the queried information, and the answer depends on the past history of recording criteria and heuristics used inside a specific sensory mode.
5. Irrationality
The M-Logic Machine does not follow the standard AI definition of a rational agent: It does not seek to maximize some performance measurement. It's only implicit goal is to exist, so performance only needs to be "good enough". In this sense, it's behaviour is often "irrational". Sub-optimal sequences of actions can be learned, and in some settings this leads to Skinners's "superstitious learning". The M-Logic Machine will tend to cooperate in an iterated prisonner's dillemma. The "keep-exploring principle" will drive the machine to leave possibly optimal solutions (like humans do in the Iowa Gambling Task).
6. Belief Incoherency
The dominance mechanism that is used to generate reliable beliefs is not logically coherent. Depending on its specific situation and experience, the M-Logic Machine may believe that birds fly, and yet believe that penguins are birds and do not fly. In the M-Logic Machine, belief is often incoherent. Only knowledge is coherent.