rowan.a.tl/propeller
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

overhaul to down-sampling. Separated selection from the down-sampling type.

Browse Source 
Also, added to the docs to help startup faster
This commit is contained in:
Ryan Boldi 2022-12-12 11:53:14 -05:00
parent b4e4552acb
commit e3ef43e95a
5 changed files with 66 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

59
doc/downsample.md
View File

@ -1,27 +1,56 @@
# Downsampling the Training Data
Downsampling the Training Data
=
Downsampling is a very simple way to improve the efficiency of your evolutionary runs. It might allow for deeper evolutionary searches and a greater success rate.
Using Downsampled-Lexicase selection with propeller is easy:
Using Downsampled selection with propeller is easy:
Set the :parent-selection argument to whichever selection strategy you would like, and set the :downsample? argument to true as follows:
Set the :parent-selection argument to :ds-lexicase as follows
```clojure
lein run -m propeller.problems.simple-regression :parent-selection :ds-lexicase <ARGS>
lein run -m propeller.problems.simple-regression :parent-selection :lexicase :downsample? true
```
Arguments:
The number of evaluations is held constant when comparing to a full training set run, so set the :max-generations to a number of generations that you would have gone to using a **full** sample.
## Downsample Functions
In this repository, you have access to 3 different downsampling functions. These are the methods used to take a down-sample from the entire training set.
To use them, add the argument ```:ds-function``` followed by which function you would like to us
The list is
- ```:case-maxmin``` - This is the method used for informed down-sampled lexicase selection
- ```:case-maxmin-auto``` - This method automatically determines the downsample size
- ```:case-rand```- Random Sampling
### Using ```:case-maxmin```:
In order to use regular informed down-sampled selection, you must specify a few things:
- ```:downsample-rate```- This is the $r$ parameter: what proportion of the full sample should be in the down-sample $\in [0,1]$
- ```:ds-parent-rate``` - This is the $\rho$ parameter: what proportion of parents are used to evaluate case distances $\in [0,1]$
- ```:ds-parent-gens``` - This is the $k$ parameter: How many generations in between parent evaluations for distances $\in \{1,2,3, \dots\}$
### Using ```:case-maxmin-auto```:
In order to use automatic informed down-sampled selection, you must specify a few things:
- ```:case-delta ```- This is the $\Delta$ parameter: How close can the farthest case be from its closest case before we stop adding to the down-sample
- ```:ids-type``` - Either ```:elite``` or ```:solved ``` - Specifies whether we are using elite/not-elite or solved/not-solved as our binary-fication of case solve vectors.
- ```:ds-parent-rate``` - This is the $\rho$ parameter: what proportion of parents are used to evaluate case distances $\in [0,1]$
- ```:ds-parent-gens``` - This is the $k$ parameter: How many generations in between parent evaluations for distances $\in \{1,2,3, \dots\}$
### Using ```:case-rand```:
In order to use regular randomly down-sampled selection, you must specify a few things:
- ```:downsample-rate```- This is the $r$ parameter: what proportion of the full sample should be in the down-sample $\in [0,1]$
- Case Downsampling function:
- Random sampling (default)
- Case tournament selection
Here's an example of running informed downsampled lexicase selection with $r=0.1$, $\rho=0.01$ and $k=100$ on the simple classification problem:
```clojure
:ds-function :case-tournament
```
- Case Lexicase Selection
WIP
- Downsample Rate:
```clojure
:downsample-rate 0.1
lein run -m propeller.problems.simple-classification :parent-selection :lexicase :downsample? true :ds-function :case-maxmin :downsample-rate 0.1 :max-generations 300 :ds-parent-rate 0.01 :ds-parent-gens 100
```

33
src/propeller/downsample.cljc
View File

@ -20,28 +20,6 @@
[training-data {:keys [downsample-rate]}]
(take (int (* downsample-rate (count training-data))) (shuffle training-data)))
(defn select-downsample-avg
"uses case-tournament selection to select a downsample that is biased to being spread out"
[training-data {:keys [downsample-rate case-t-size]}]
(let [shuffled-cases (shuffle training-data)
goal-size (int (* downsample-rate (count training-data)))]
(loop [new-downsample (conj [] (first shuffled-cases))
cases-to-pick-from (rest shuffled-cases)]
;(prn {:new-downsample new-downsample :cases-to-pick-from cases-to-pick-from})
(if (>= (count new-downsample) goal-size)
new-downsample
(let [tournament (take case-t-size cases-to-pick-from)
rest-of-cases (drop case-t-size cases-to-pick-from)
case-distances (metrics/mean-of-colls
(map (fn [distance-list]
(utils/filter-by-index distance-list (map #(:index %) tournament)))
(map #(:distances %) new-downsample)))
selected-case-index (metrics/argmax case-distances)]
(prn {:avg-case-distances case-distances :selected-case-index selected-case-index})
(recur (conj new-downsample (nth tournament selected-case-index))
(shuffle (concat (utils/drop-nth selected-case-index tournament)
rest-of-cases))))))))
(defn select-downsample-maxmin
"selects a downsample that has it's cases maximally far away by sequentially
adding cases to the downsample that have their closest case maximally far away"
@ -58,10 +36,6 @@
(utils/filter-by-index distance-list (map #(:index %) tournament)))
(map #(:distances %) new-downsample)))
selected-case-index (metrics/argmax min-case-distances)]
(if (sequential? (:input1 (first new-downsample)))
(prn {:cases-in-ds (map #(first (:input1 %)) new-downsample) :cases-in-tourn (map #(first (:input1 %)) tournament)})
(prn {:cases-in-ds (map #(:input1 %) new-downsample) :cases-in-tourn (map #(:input1 %) tournament)}))
(prn {:min-case-distances min-case-distances :selected-case-index selected-case-index})
(recur (conj new-downsample (nth tournament selected-case-index))
(shuffle (utils/drop-nth selected-case-index tournament))))))))
@ -81,13 +55,8 @@
selected-case-index (metrics/argmax min-case-distances)]
(if (or (= 0 (count tournament)) (<= (apply max min-case-distances) case-delta))
new-downsample
(do
(if (sequential? (:input1 (first new-downsample)))
(prn {:cases-in-ds (map #(first (:input1 %)) new-downsample) :cases-in-tourn (map #(first (:input1 %)) tournament)})
(prn {:cases-in-ds (map #(:input1 %) new-downsample) :cases-in-tourn (map #(:input1 %) tournament)}))
;(prn {:min-case-distances min-case-distances :selected-case-index selected-case-index})
(recur (conj new-downsample (nth tournament selected-case-index))
(shuffle (utils/drop-nth selected-case-index tournament)))))))))
(shuffle (utils/drop-nth selected-case-index tournament))))))))
(defn get-distance-between-cases
"returns the distance between two cases given a list of individual error vectors, and the index these

28
src/propeller/gp.cljc
View File

@ -17,13 +17,14 @@
(defn report
"Reports information each generation."
[evaluations pop generation argmap]
[evaluations pop generation argmap training-data]
(let [best (first pop)]
(clojure.pprint/pprint {:generation generation
:best-plushy (:plushy best)
:best-program (genome/plushy->push (:plushy best) argmap)
:best-total-error (:total-error best)
:evaluations evaluations
:ds-indices (map #(:index %) training-data)
:best-errors (:errors best)
:best-behaviors (:behaviors best)
:genotypic-diversity (float (/ (count (distinct (map :plushy pop))) (count pop)))
@ -35,12 +36,13 @@
(defn gp
"Main GP loop."
[{:keys [population-size max-generations error-function instructions
max-initial-plushy-size solution-error-threshold mapper ds-parent-rate ds-parent-gens dont-end ids-type]
max-initial-plushy-size solution-error-threshold mapper ds-parent-rate ds-parent-gens dont-end ids-type downsample?]
:or {solution-error-threshold 0.0
dont-end false
ds-parent-rate 0
ds-parent-gens 1
ids-type :solved ; :solved or :elite
downsample? false
;; The `mapper` will perform a `map`-like operation to apply a function to every individual
;; in the population. The default is `map` but other options include `mapv`, or `pmap`.
mapper #?(:clj pmap :cljs map)}
@ -55,12 +57,12 @@
(fn [_] {:plushy (genome/make-random-plushy instructions max-initial-plushy-size)})
(range population-size))
indexed-training-data (downsample/assign-indices-to-data (downsample/initialize-case-distances argmap))]
(let [training-data (if (= (:parent-selection argmap) :ds-lexicase)
(let [training-data (if downsample?
(case (:ds-function argmap)
:case-avg (downsample/select-downsample-avg indexed-training-data argmap)
:case-maxmin (downsample/select-downsample-maxmin indexed-training-data argmap)
:case-maxmin-auto (downsample/select-downsample-maxmin-adaptive indexed-training-data argmap)
(downsample/select-downsample-random indexed-training-data argmap))
:case-rand (downsample/select-downsample-random indexed-training-data argmap)
(do (prn {:error "Invalid Downsample Function"}) (downsample/select-downsample-random indexed-training-data argmap)))
indexed-training-data) ;defaults to random
parent-reps (if (zero? (mod generation ds-parent-gens)) ;every ds-parent-gens generations
(take (* ds-parent-rate (count population)) (shuffle population))
@ -74,21 +76,17 @@
(partial error-function argmap training-data)
population))
best-individual (first ds-evaluated-pop)
best-individual-passes-ds (and (= (:parent-selection argmap) :ds-lexicase) (<= (:total-error best-individual) solution-error-threshold))]
(prn {:ds-indices-list (map #(:index %) training-data)})
;(if (sequential? (:input1 (first training-data)))
;(prn {:ds-inputs (map #(first (:input1 %)) training-data)})
;(prn {:ds-inputs (map #(:input1 %) training-data)}))
best-individual-passes-ds (and downsample? (<= (:total-error best-individual) solution-error-threshold))]
(if (:custom-report argmap)
((:custom-report argmap) evaluations ds-evaluated-pop generation argmap)
(report evaluations ds-evaluated-pop generation argmap))
(report evaluations ds-evaluated-pop generation argmap training-data))
;;did the indvidual pass all cases in ds?
(when best-individual-passes-ds
(prn {:semi-success-generation generation}))
(cond
;; Success on training cases is verified on testing cases
(if (or (and best-individual-passes-ds (<= (:total-error (error-function argmap indexed-training-data best-individual)) solution-error-threshold))
(and (not= (:parent-selection argmap) :ds-lexicase)
(and (not downsample?)
(<= (:total-error best-individual) solution-error-threshold)))
(do (prn {:success-generation generation})
(prn {:total-test-error
@ -100,10 +98,10 @@
false)
nil
;;
(and (not= (:ds-function argmap) :case-maxmin-auto) (>= generation max-generations))
(and (not downsample?) (>= generation max-generations))
nil
;;
(and (= (:ds-function argmap) :case-maxmin-auto) (>= evaluations (* max-generations population-size (count indexed-training-data))))
(and downsample? (>= evaluations (* max-generations population-size (count indexed-training-data))))
nil
;;
:else (recur (inc generation)
@ -117,7 +115,7 @@
(first reindexed-pop)))
(hyperselection/log-hyperselection-and-ret (repeatedly population-size ;need to count occurance of each parent, and reset IDs
#(variation/new-individual reindexed-pop argmap)))))
(if (= (:parent-selection argmap) :ds-lexicase)
(if downsample?
(if (zero? (mod generation ds-parent-gens))
(downsample/update-case-distances rep-evaluated-pop indexed-training-data indexed-training-data ids-type) ; update distances every ds-parent-gens generations
indexed-training-data)

2
src/propeller/problems/simple_classification.cljc
View File

@ -88,11 +88,13 @@
:case-t-size (count (:train train-and-test-data))
:ds-parent-rate 0
:ds-parent-gens 1
:ds-function :case-rand
:max-generations 500
:population-size 500
:max-initial-plushy-size 100
:step-limit 200
:parent-selection :lexicase
:downsample? false
:tournament-size 5
:umad-rate 0.1
:variation {:umad 1.0 :crossover 0.0}

3
src/propeller/selection.cljc
View File

@ -26,5 +26,4 @@
[pop argmap]
(case (:parent-selection argmap)
:tournament (tournament-selection pop argmap)
:lexicase (lexicase-selection pop argmap)
:ds-lexicase (lexicase-selection pop argmap)))
:lexicase (lexicase-selection pop argmap)))