23rd Feb 1995
When executing a sequence from experience that has been successful in the past there is no need to store the new sequence because it is the same as the past. It serves no purpose. Only need to record it in LTM if something new occurs. It would exist in STM and not be transferred to LTM unless something new occurs and then it would be saved Long Term. A mark would be placed in LTM to identify where the missing piece was and no associations could be made across the marker.
A modification to the pattern matching of a sequence (series) of stimulus / responses over time to obtain the best match and provide the best context to the recognition and recall process and where to execute is to allow stimuli to match over a time sequence independent of the responses matching in a sequence. See 10th Feb 1985.
Thinking is more than just repeating the Stimulus-Response experiences from memory. It must include a decision making process based on feelings and a strategy for finding better responses than the punished ones that may be recalled. Do we continue to mentally search for (recall) a response that is neutral or positive by repeating the cue?
Success and failure are synonymous with reward and punishment. So in a learning situation learning is taking place when behaviour is modified so that success is achieved.
See the item in STM subject area of this same date for a mechanism associated with parallel activity that is not recorded long term.
The next simulation to try is one in which the S/R/F are stored as before but so are the recalled S/R/F stored. However the recalled ones are not used for associations or recall. We have to make sure the S/R/F simulation still works and that the recall / association process skips over the recalled S/R/F recordings.
12th Apr 1995
I stored all recalled S/R/F but did not recall them or match them up. The simulation worked identical to the above. I then changed the pattern matching for recall so that a sequence of inputs matches over time independent of the interspersed outputs and feelings or ideas of S/R/F. The only change was that the repetitive negative recall loop was broken because an earlier recall (not punished) was longer now that intermediate R's were skipped and it carried on successfully. However a newer negative recall loop arose in the 9 cell environment.
The second idea from 11th Nov 94 is that the response should be recalled before it is output ideally to find any feelings associated with the response anywhere else in memory other than the one about to be executed. My first thought is the feeling associated with the response at the "about to be" execution point is the one that is relevant as a gauge of the association strength between stimulus and response, not other feelings associated with this response in other contexts. Second, feelings are internally generated inputs and the feeling associated with the response should be generated at the time the response is recalled by the association lookup process. An example in real life is that I reach out and touch many things during my life but I do not stop from reaching out and touching because at one time in my past I reached out and touched fire and got burnt. My reach out and touch response has been learnt in the context of the stimulus (fire or cookie). This means that responses are not inherently bad in themselves, only in the context of a stimulus.
You cannot get bored with repeating the same sequence of outputs. It’s only inputs that can be boring.
25th Aug 1995
It was said at the IJCAI conference that the Turing Test is an imitation game and the real objective is to invent software which magnifies human intelligence. However I believe it is to understand the principles of intelligent behaviour in order to duplicate it. Thus create self aware, learning and thinking behaviour which can be placed in all manner of things which will help man achieve his goals.
Try modelling Adaptron into pieces that correspond to what the AI community is trying to optimise, i.e. induction, decision making, search etc. The Adaptron architecture may be more open to further research and the application of existing results.
Forgetting is important - Adaptron - its old info. - Unused - is less likely to be recalled - used.
We have the ability to build concepts of combinations of concepts and deal with them like any other concept - to any order of hierarchy / complexity - Does language play a part in this? We have the ability to name things, which is similar to identifying complex concepts.
I need to analyze how we create sounds that are the same as the sounds we have heard in the past - The imitation learning process.
Thinking consists of:
Induction - concept forming from examples - recognition in thought / learning (learning of concepts).
Deduction - reasoning from facts to form a conclusion
Problem Solving = decision making (deduction?)
Choose a sequence to execute = decision making - at each step in the execution should have the opportunity to change to another sequence which better fits sensory inputs / feelings, etc. This would occur when a recall interrupts the execution because its feeling level is equal or higher than the current concentration level.
Do we have a concept of a goal before we start any action? We do have a concept of need to achieve a feeling that was stored when recalled from association with the cue. Goal would be the associated input and its feeling. If goal still in STM when achieved - store sequence in LTM! The decision to start executing is based on the idea that you will achieve the goal.
Some examples of Good / Neutral / Bad feelings are:
Have achieved goal
Can't obtain any recall matching attention wanted.
Do feelings play a part in deciding what gets stored in Long Term Memory? Probably!
Feeling about - make a mistake! Having recognized boredom, mistake etc. then generates the feeling that is attended to, memorized and recalled - associated.
Why should it do anything? Given an environment, Good & Bad it needs to have motivation, a need to satisfy. Goals - reach them and the satisfaction (which is a feeling) decays so that you need to pursue them again else nothing happens after goal reached.
Marvin Minsky talked about pain interrupting all high level learnt behaviour and all the hierarchy of goals that you have formulated over your lifetime.
Intelligent agents need not talk to exhibit their intelligence. Deaf people use sign language and are still seen as being intelligent. Just as long as behavioural symbols match (map to) the internal thought concepts then communication is possible.
Communication uses a "sensory" domain specific for the purpose of communication - Language.
There is a need for a test to recognize and acknowledge that something is learning /thinking /sentient. It needs to test for the following aspects /dimensions /processes of intelligence: Recognition, Learning, Thinking, and Communicating. Then there is a need for a systems test which tests all of these working together to exhibit intelligent behaviour.
Environments can be divided according to the following criteria:
Static vs Dynamic
Deterministic vs Non-deterministic
Completely Detectable vs Partially Detectable
Predictable vs Non-predictable
Continuous vs Discrete