From d341a2bbad41fd1def2bf46806fc572744b66533 Mon Sep 17 00:00:00 2001 From: olpc user Date: Sat, 7 Dec 2019 10:33:25 -0800 Subject: notes on learning with this much basis --- .../habit-starts/learn-to-dance-level-1.txt | 89 ++++++++++++++++++++++ 1 file changed, 89 insertions(+) create mode 100644 starts/meaning-vm/habit-starts/learn-to-dance-level-1.txt (limited to 'starts/meaning-vm') diff --git a/starts/meaning-vm/habit-starts/learn-to-dance-level-1.txt b/starts/meaning-vm/habit-starts/learn-to-dance-level-1.txt new file mode 100644 index 0000000..cad85f3 --- /dev/null +++ b/starts/meaning-vm/habit-starts/learn-to-dance-level-1.txt @@ -0,0 +1,89 @@ +'validate your reason for existence' relates directly to pattern learning. + +The validation is a pattern of what is good about us, most simplistically +a reference to a past event we were involved in, where we contributed/succeeded. +Preferably a pattern of us being able to reproduce good events. + +This is a way to learn to dance. + +Say we have a habit that has an unknown delay before firing, and we want to fire +it in synchrony with an event. Our goal is to produce our event within a smaller +time window to the target event than in the past ("same time as"). +Good: time window is closer than ever before. + +need: history log to refer to good event. + please wait a little? expanding reference to good event into how-to-learn + need: behavior based on what-succeeded, what-failed + value metric? + +SO! we want to learn how to time an event. We have some tools, for example: + - waiting for or until a given time + - getting the current time + - comparing two times +We want to combine the tools in a way that makes the event happen at the time +we want. + - doing something after the right time happens + - doing our event +Since each habit has an unknown delay, we might play with delaying a certain +time since the last event, until we find the right delay that works best for us +most of the time. + Testing metric: runs when event is fired, measures time between + event and right time. if time is less than ever before, success. + if time is significantly more than behavior's norm, failure. + Convert to English: try to have the event happen at the right time. + note metric will give random successes false status + +A successful approach would be to adjust the delay towards the difference by +a small ratio. +The most successful approach would be to use the time difference to adjust the +delay precisely. + Ideally we would find solution #2 after trying solution #1. + The idea of 'moving towards' would adjust into 'moving the exact right + amount'. + In operators, this could be a development of the subtraction operator. + But using a value exactly is actually simpler than using a ratio of it. + So we can move from numbers towards ideas. + More. Less. More a lot? Less a lot? More a little? Less a little? + Ideally we use learning strategies that facilitiate learning + how to learn in general. + That means summarizing and acting on the meaning of pattern structures. +In reality, everything jitters a little bit. Nothing is ever exactly the same. +Things also grow and shrink over time. + +Habits look to be needed, to have value. +As one ourselves, we look to relate to those that meet our needs, have value to +us. +The core habit, to learn, is the one that selects other habits and works with +them. Ideally it's an intermixing of existing habits. + +What might a winning habit's structure look like? say it is the perfect one. +set to do 1ce on goal time: + ctx X + record time A + set to do 1ce on goal time: + record time B + trigger D1 for X + delay for C (X) + record time E + trigger D2 for X + When both D1 and D2 have been triggered for X: + calculate B - E, store in F + calculate F + C, provide as C for next context + + will want to know which C is being used when we delay. + could be wrong C. + + and we'll want to form structure promises ... + .. and map to meaning for operator + operator watches and understands as learning + develops, and provides labels for shared + understanding that develops. + operator will want generalization to happen + fast, so as to label shared meaning. + could also provide label-goals, and code must guess + towards goals, to get onto same page as operator. + +I think in structuring such a large habit out of parts, we would find a lot +of learning relevence. + + -- cgit v1.2.3