PereGaea
Laslo Godel
DECISIONS
The Dynism clearly must become more than an increasingly complex machine-like automaton reacting to stimuli with `hard-wired' responses. We will now look at how it might acquire the ability to make `Decisions' and, as a result, become able to learn from experience and begin to behave as if it possessed `consciousness'. We can of course never know absolutely if it will ever possess consciousness, let alone allow us an insight into it.
To begin, what happens when the Dynism encounters two Objects, Predators for example, that it must respond to simultaneously? Which does it respond to first? And just to make things more complicated, what if the objects are different from each other, are at different distances, have different orientations, configurations, or are on the move? And, of course, what happens if there are more than two of them?
The simplest mechanism the Dynism can evolve is, as we've already seen, one that decides between all these things at random. But this is somewhat hit and miss. Wouldn't it be better if the Dynism could somehow evolve `predecisions' in its ROM for as many combinations of objects and attributes as it can through the experience of Evolution? Unfortunately, even if its anatomy could rearrange itself like this, it wouldn't be able to store more than a very few of the vast numbers possible.
The Dynism now evolves a better mechanism that uses the internal state sensors and their single-place Gain Loss flags we saw when we looked at Signals. To show you how it works, let's assume this `Decider' initially only allows the Dynism to Decide between two Significant Objects. These must have different Attributes so that the Decider can distinguish between them, otherwise it can still only select at random.
If this condition is met, the Decider then places their Object Numbers into a piece of RAM as a `Pair' like this
The Decider now selects one object at random and causes the Dynism to respond to it first. If the outcome of doing so after a `brief' interval (determined via the experience of evolution) is a Gain of some kind, a Gain Flag is attached to that object's Number in the Pair.
When the Dynism encounters the next pair of Significant Objects, its Decider automatically checks their Numbers against the Pair in the Buffer. If they match, the Dynism then responds first to the Object whose Number has the Gain Flag. Had the first Decision resulted in a Loss and a Loss Flag attached instead, the
Dynism would instead respond first to the other object in the Pair.
The same Flags, either Gain or Loss, may come to be attached to both Objects when a Decision situation is re-encountered. The Dynism can only Decide between them at random from then on unless the objects finally do acquire different flags. Clearly, the longer the Dynism lives, the greater the number of Number Pairs that will need to stored in the Decision RAM. Eventually it will completely fill. Relatively few of the Object Pairs stored in it however will remain useful forever; many Decisions may only need to be made once in a lifetime. They are therefore Datestamped. Once the RAM fills, those pairs that have not been matched by that time are simply overwritten with new ones.
More advanced Dynisms may now evolve Decision Flags for four levels of Gain or Loss. These can then be associated with Datestamps with four different durations. If the Dynism makes a Decision that results in a Gain or Loss four times greater than the minimum, the Object Pair that gave rise to it acquires a Datestamp that will survive four memory reclamation cycles rather than just one. Decisions resulting in large Gains or Losses will therefore tend to be `remembered' longest.
In this way, Decision begins to evolve into learning by experience. Natural Selection, just as with the evolution of Dynisms themselves, is now alive and well within their memory systems. If a Pair is useful, it continues to live on through successive regenerations, if it doesn't, it `dies'. The only aspect of the RAM that is controlled by `corporeal' evolution is its length. This must adjust itself so that the number of permanent Pairs does not grow to the point where new Pairs cannot be stored long enough to prove their usefulness. Its structure will also need to have the same parallel access structure as the ROM to allow pairs to be matched quickly.
Let's now look at a slightly different aspect of Decisions by coming back to another question I raised earlier: how do Dynisms Decide when and when not to perform particular Actions?Just as Dynisms acquired the ability to identify ROM Actions and Sequences, they now acquire the ability to recognize RAM Action Sequences. To do so, they must begin by inserting all the Actions they perceive, whether they perform them themselves or observe others to, including Signals, in a continuous stream into their RAMS. This has a slightly different structure from the Decision RAM in that, when it fills, the oldest Sequences are simply deleted. A new mechanism we will call a `Correlation Tester' constantly compares each new Sequence with previous ones to find a nearest-match. The start and end Configs of each such Sequence are determined solely by those of any previously stored Sequence they are matched to; this becomes the Template. Here I've drawn individual Actions using the symbols we saw before and the RAM Sequences they form part of as rectangles of various colors:
Sequences need not match each other exactly in the way I've shown here, other Actions can be interspersed. These interspersed Actions must however be followed by at least three from the Sequence they interrupt so that the Correlation Tester will treat them as an interruption, not as the beginning of a new Sequence. The interruption can itself be matched to another Sequence; this may happen for instance if the Dynism must go round some obstacle while it is pursuing a Prey:
The Correlation Tester soon becomes more than a simple Nearest-Matcher however. It first matches a Sequence to a Template solely with respect to its Actions, then attaches Attributes to it that represent certain situational factors. These may include such things as the speed at which the Sequence is performed, the object a component Action is performed with if it is Transitive, the Dynamic Object performing the Action, and its Individual Identity if this is a Dynism.
In these examples in which the Actions all match, Sequence 1 also matches only with respect to the Speed at which it is performed. In Sequence 2 only the Transitive Object Attribute matches, With Sequence 3, this is the only Attribute that doesn't match.
Recognizing RAM Action Sequences is, like many new capabilities the Dynism evolves, all but useless until it also acquires additional mechanisms. We'll call the first of these an `Assessor'. Along with all the Actions and Signals a Sequence can contain, it can now also include Gain or Loss Flags, whether the Gains or Losses that give rise to them result from Decisions or not. What this means at this stage is that, if the Sequence a Dynism performs in response to a Significant Object begins to match one that ended in a Loss, the Dynism will almost certainly eventually undergo that same Loss if it continues to perform it.
Its Assessor however gives the Dynism two ways of avoiding this fate. The `Object' Decisions we saw earlier are now also placed into a Sequence as they occur rather than being stored in their own piece of RAM as before. The `brief' interval before a Gain or Loss Flag can be attached to it is removed, these too can now be attached to a Sequence at any point along its length, If the Dynism is currently re-enacting a Sequence containing a Loss Flag, the Assessor looks along it for a record of a Decision that the Dynism made before the Loss occurred. It will then cause the Dynism to takes its alternative option once it reaches that point. The Loss may then be avoided - indeed the Dynism may even perform a new Sequence that ends in a Gain Flag.
If no such Decision point exists, the Assessor may cause the Dynism to select a Sequence that also nearest-matches the one it is performing, but has different Attributes. If this involves different Transitive Objects, the Dynism must then use its Targeting Reflexes to accommodate these. If the alternative Sequence contains a Gain Flag, there is no guarantee the Gain will be repeated, indeed the Sequence may even produce a Loss. But if it does succeed, the Dynism not only acquires its Gain, but it now has another Sequence it can use in the same way next time it finds itself in a similar situation. In this way the Dynism can cope with new experiences by adapting previous ones. Although this will not always be successful, it should at least reduce the risk of disaster.
The Correlators in slightly more advanced Dynisms may periodically examine all such nearest-matching Sequences in its RAM to determine whether a particular Attribute made any difference to their outcomes. If not, this Attribute can be replaced in all of them with a Null Attribute meaning that, should they be matched again, this Attribute can be ignored. In other words, any such Sequence may be re-enacted irrespective of any Transitive Object it may be performed on or, perhaps, by any Dynism rather than a particular one.
With their multilevel Gain-Loss Flags, Dynisms may now become able to `make sacrifices' in order to pursue higher Gains. If the Dynism finds itself re-enacting a Sequence containing a low-level Loss followed by a high-level Gain, the Dynism will continue to re-enact that Sequence. It will also do so if the situation is reversed; that is, the Loss follows the Gain. It will not however re-enact a Sequence where both are at low levels.
The Assessor may also become able to take Time into account with such `Plus and Minus' Sequences. The longer the time-interval between a Loss and a following Gain for instance, the greater the chances that the Dynism's situation will change and prevent the Gain from coming about. The Assessor now attaches a `Score' Flag to each such Sequence which records the Gains and Losses after several re-enactments. If overall Gains outweigh Losses the Dynism will continue to re-enact it, otherwise it will always `jump' to another via its Decider in the way we saw earlier.
Its worth noting that, while Sequences may be recorded in the RAM in sequential order, the Dynism certainly need not re-enact them that way. For instance, as we've just seen, re-enacting a Sequence containing a Loss will cause a Dynism to jump to another containing a Gain. This second Sequence may in fact have several such `pre-entry' Sequences leading to it, each of which may have pre-entry Sequences leading to them. Any one of these therefore can place the Dynism in a situation where it can follow a `primeval' Sequence to attain a Gain. In this way such Sequences `connect' to each other as the tributaries of a river do, indeed one that spreads throughout its RAM as a Dynism ages and acquires experience.
