Re:Engineering Kindness: Building A Machine With Compassiona...
by Animats on Saturday March 04, 2017 @04:05PM
(#13792926)
I wasn't expecting compassionate intelligence via predicate calculus based expert systems. That's very 1980s, as is the whole paper. There's no mention of machine learning. There's a lot here about "common sense", which remains an unsolved problem in AI.
Thinking about common sense without getting into philosophical tangles is difficult. It's helpful to consider common sense as the ability to predict the consequences of actions.
If you can predict consequences, (probably probabilistically) you have the basis for planning. And you can avoid getting into trouble, which is what common sense is mostly for.
This is known technology. In modern control theory, you set up a predictor, have it watch the "plant", and self-tune until it can predict what the plant will do. This is called adaptive control, and it's a kind of machine learning, with similar math but different terminology. Once you have a good predictor, you can invert the predictor function and solve for the inputs which will yield the desired outputs. This works on systems which are reasonably continuous. Many industrial processes are controlled by such systems.
Prediction is a prerequisite for compassionate intelligence. It's not having internal emotional state that's crucial. It's having a predictive model of the other party's emotional state.
Of course, once you have a tuned model of somebody, you now know what buttons to push to make them do things, in either the "buy" or "vote" spaces. Whether something is compassionate is a design decision.
Thinking about common sense without getting into philosophical tangles is difficult. It's helpful to consider common sense as the ability to predict the consequences of actions. If you can predict consequences, (probably probabilistically) you have the basis for planning. And you can avoid getting into trouble, which is what common sense is mostly for.
This is known technology. In modern control theory, you set up a predictor, have it watch the "plant", and self-tune until it can predict what the plant will do. This is called adaptive control, and it's a kind of machine learning, with similar math but different terminology. Once you have a good predictor, you can invert the predictor function and solve for the inputs which will yield the desired outputs. This works on systems which are reasonably continuous. Many industrial processes are controlled by such systems.
Prediction is a prerequisite for compassionate intelligence. It's not having internal emotional state that's crucial. It's having a predictive model of the other party's emotional state.
Of course, once you have a tuned model of somebody, you now know what buttons to push to make them do things, in either the "buy" or "vote" spaces. Whether something is compassionate is a design decision.