Old-fashioned (OpenNARS) idea: Anticipation with explicit timeout

[patham9]

Via postdiction: by only looking at belief event and predicted belief in the concepts some time after the timeout has passed (looking whether there was evidence from input within the predicted timeframe). This will likely not be merged, unless it turns out to be better, but so far it seems worse, a key reason:

• With old-fashion anticipation it can happen that there will be missed opportunities to notice failure of a hypothesis.
• With assumption of failure based anticipation however, this can never happen (since a piece of neg. evidence is added right away when the hypothesis is used). it can only happen to punish too much, never too little
• Timing estimation (including time tolerance estimation) needs to be quite accurate for classical Anticipation to work well, while Assumption of Failure does not depend on the actual timing estimates. Also evidence counting will be inconsistent if time deltas are adjusted, which can only be circumvented by leaving hypotheses with quite different time deltas separate rather than merging them (resource issue, and complexity in deciding when they are similar enough to merge)

[PtrMan]

I would call it simply as anticipation instead of "old fashioned anticipation", to me "assumption of failure" isn't anticipation because nothing is anticipated in a finite time window.

Also evidence counting will be inconsistent if time deltas are adjusted

Non existing evidence shouldn't be evidence in the first place. With anticipation no revision takes place with virtual evidence. The issue doesn't exist if there is no revision of virtual evidence.

[patham9]

"I would call it simply as anticipation instead of "old fashioned anticipation", to me "assumption of failure" isn't anticipation because nothing is anticipated in a finite time window."

That's not true. Assumption of failure needs the event to happen within 20 steps or so. Else truth expectation will go down rather than up. So it's an implicit time window of 20 steps which is at the same time also the temporal window size the system can find positive evidence for without exploiting transitivity via sub-goaling. Also within the 20 steps, the sooner, the more impact the positive evidence will have, instead of a binary decision, which is harder to realize with old-fashioned anticipation, even though we tried it (though poorly) in the Clojure version.

"Non existing evidence shouldn't be evidence in the first place. With anticipation no revision takes place with virtual evidence. The issue doesn't exist if there is no revision of virtual evidence."

Not sure what you are talking about. Clearly if dt=5 and then after revision dt=10, classical anticipation will assign negative evidence differently after revision, unless the hypotheses with dt=5 and dt=10 are kept separate.