294 William St.
4th Dec. 1973
Professor A. Newell,
I will be graduating with an M.Sc. in Computer Science from Queen’s University in May 1974. I am extremely interested in doing post-graduate research in Artificial Intelligence at Carnegie-Mellon University.
I have been working independently on a simulation of human learning and thinking processes. I would be very glad if you would give your opinion as to whether or not the enclosed ideas are worth while pursuing at the Ph.D. level.
Although the description is very general and ill-defined, I have specified a major part of it in the form of a PDP 11/45 assembler program (approximately 500 instructions). I have taught it stimulus-response type sequences. There are many areas in which the simulation can be improved, such as the attention interrupt structure, and the short-term memory processes.
Thank you for your kind attention.
Brett N. Martensen
A Brief Description of the simulation of Human Learning and Thinking processes.
There is an input device (teletype) and an output device (the same teletype). All input and output is stored as a linear sequence of records in the memory. An ‘attention’ process selects possible information to be recorded. It can select from the following sources, Input, Short-term memory and association from long-term memory. The three sources can be thought of as input channels. Any associations from Long-term memory which are ‘paid attention to’ are recorded as ‘concepts’. As every record is stored in memory a recall process takes place, going sequentially back through memory finding similar occurrences of the record. At these points associated inputs, outputs and concepts are found. What is meant by an association has been explicitly defined by the program.
There are two input symbols which have predefined meaning to the system. They represent punishment and reward. Another two symbols also have meaning to the system. These are the concept of punishment and the concept of reward. The selection of the appropriate information by the attention is guided by the ‘strength’ of the information. Information which is associated with reward is ‘stronger’ than information associated with the concept of reward etc.
The system can be in either of two states.
1/ Waiting for input and recording information which attracts attention due to the recall process.
2/ Paying attention to the same sequence of input channels as was done on a previous occasion. In this state of ‘execution’ only very ‘strong’ information can interrupt.
To get into this state of ‘execution’ the system has to have an input which is associated with reward or the concept of reward. If punishment or the idea of punishment attracts attention the simulation will change to state one.
The system starts in state one and memory is empty, except for a prerecorded instinct to echo the second symbol of a two symbol sequence. This is used to initiate responses. The system also has a reflex, which is to output a symbol if attention is attracted to that symbol because it was rewarded on a past occasion.
B N Martensen
3rd January 1973 The reply from Allen Newell