Time for a plan.

I don’t have Social Salience, so I treat other humans as Black Boxes with some attached facts. I may know their stated preferences, stated ideas, etc. Given a goal I can construct a logical model of the situation where I can compute the inputs that I think will give the desired output(s).

This involves:

• retrieving facts and opinions from associative memory
• computing expected value for uncertain variables
  • OK, I am not literally doing math here but estimating something like P(50%) * HIGH > P(90%) * LOW
• more Propositional Logic to model everything

Obviously this is expensive, perhaps more so than Theory-Theory. Luckily I have developed some optimizations:

• familiar situations, e.g. going to lunch, have a script (zero cost)
  • plug in the people present
  • proceed
• low fidelity models can be used for many situations (low cost)
  • if the cost of mistakes are low use a low fidelity model
    • e.g. other situations that had some similarity
    • not considering all the people and factors involved: usually a logical argument will suffice
  • own the cost of a mistake – accept that they can happen
  • understand that the data is incomplete
    • use risk management strategies if needed
    • make a conscious decision to proceed
  • adjust via feedback
  • typical random situations
• high fidelity models can be used when needed (high cost)
  • for example, a presentation to another group to convince them to do a certain project
  • consider the available data, logical arguments, how people may react, what are their likely needs and desires given what their group does, etc.
  • I need to model complex black boxes that represent people with enough fidelity that I can predict which inputs will produce the desired outputs
    • without the input of vibes or other social salience
    • just data and logical arguments
    • person A’s stated goals
    • known constraints of the environment
    • logical causal chains (“Given Person A’s stated facts and constraints, a logical agent would produce Y if given input X.”)
  • this is not dissimilar to NT people planning the same meeting, just focusing on pure information and logic
    • NT people need to consider social and emotional aspects as well
  • this is not used real-time, this is planning for a meeting or other even and asynchronous

Feedback Loop

There is one additional heuristic:

• replan if necessary (low to medium cost)
  • if new information is presented or I detect the current plan failing
  • come up with a new focused low fidelity plan
    • not nearly as fast as NT social signals (double empathy)
    • but usable with some computational lag in a conversation

See Typical Example for a replan – it can be done during a conversation if simple enough.

Limitations

And some disadvantages:

• the replan trigger requires explicit information
  • if the NT people around me have decided on a new, not explicitly stated, plan I will still be on the old plan
  • in social situations this might mean unstructured shopping time when I think it is time to go to the next destination
  • this is a significant source of social Friction with NT people – I appear stubborn or stuck
• if the task requires an emotional plea I likely cannot accomplish it
  • I can represent my own emotions but I have no affect empathy and cannot adjust

Next Step: Propositional Logic (Runtime)

The model produces a plan for how to achieve my goals – the instructions I need to execute to achieve the desired result via Propositional Logic. Those are the instructions I need to evaluate in the next step – some small subset of my available instructions.