Allow predicted evidence to contribute to observed evidence similar as originally envisioned but with a configurable parameter which balances observation and prediction similar as in Kalman filters.
This will likely not be merged until clear uses emerge.
A question which underlies this: in which cases do percepts need to be "made-up"? Isn't building stable temporal implications which allow for various ways of subgoaling (ending at realized goals corresponding to from current observations derivable evidence) even when particular percepts are absent sufficient (or even the way to go) to handle this? How is the system supposed to take into account properly the case where some predicted percepts are absent when it just proceeds as-predicted (hypothesis failure versus sensor unreliability)? Even if a tradeoff is possible, sensors should be quite trustable.
automenta
commented
2 years ago
Allow predicted evidence to contribute to observed evidence similar as originally envisioned but with a configurable parameter which balances observation and prediction similar as in Kalman filters.
This will likely not be merged until clear uses emerge. A question which underlies this: in which cases do percepts need to be "made-up"? Isn't building stable temporal implications which allow for various ways of subgoaling (ending at realized goals corresponding to from current observations derivable evidence) even when particular percepts are absent sufficient (or even the way to go) to handle this? How is the system supposed to take into account properly the case where some predicted percepts are absent when it just proceeds as-predicted (hypothesis failure versus sensor unreliability)? Even if a tradeoff is possible, sensors should be quite trustable.