Wolfram Cloud Document

Generate Problems

In[]:=

x_

p__

:=x〚p〛

In[]:=

pr=PlotRange{{0,4},{0,4}};lo={1,1};hi={3,3};gen:=RandomInstance[GeometricScene[{a,b,c},{CircleThrough[{a,b,c}],Point[a]∈Rectangle[lo,hi],Point[b]∈Rectangle[lo,hi],Point[c]∈Rectangle[lo,hi]}]]z=genρ=

z

1,1,;;3

All,2

r=Rectangle@@

pr

2,1,1

,

pr

2,2,1

,

pr

2,1,2

,

pr

2,2,2

;i=3(*activepoint*);

Out[]=

{{2.29487,1.40843},{1.58084,1.46154},{1.32209,2.76856}}

Define DeviceClass Functions

In[]:=

state=s="ObservedState";ended=e="Ended";reward="Reward";points=3;actions={Left,Right,Up,Down};f=False;t={2,2};l={0,4};finished:=Withe=f||

ρ

i

~ContainsAny~l,If[e,f=e];erandom:=RandomChoice@actionsreset:=f=False;z=gen;ρ=

z

1,1,;;points

All,2

;best=t~EuclideanDistance~ρi;readread:=<|staterender,endedfinished|>render:=ImageRotateImageSparseArrayClipRound10

ρ

i

,{1,64}1,40,0,90°Δ=.1;Δσ=<|Left{-Δ,0},Right{+Δ,0},Down{0,-Δ},Up{0,+Δ}|>;

In[]:=

step//ClearAllstep[a_]:=Module[{o,n,β,e,r},o=ρ〚i〛;n=If[f,o,ρ〚i〛=Clip[o+Δσ[a],l]];β=t~EuclideanDistance~n;e=finished;r=If[e,0,If[best≤β,0,best=β;1]];<|staterender,endede,rewardr,"Best"best|>]

Test Human interactions

In[]:=

joystick:=Grid[({{,Button["↑",last=step@Up],},{Button["←",last=step@Left],Button["",last=reset],Button["→",last=step@Right]},{,Button["↓",last=step@Down],}}),Spacings{0,0}]

In[]:=

last=reset;(*←→↑↓*)joystick{last//Dataset,best}//Column//Dynamic

Out[]=

	↑
←		→
	↓

Out[]=

last

best

Define Neural Networks for Reinforcement Learning

from Wolfram Mathematica website for CartPole demo


Loss Function

In[]:=

NeuralNetworks`Private`$ExpandNetSummaryBoxByDefault=True;netSize=Dataset@NetInformation[#,{"ArraysElementCounts","ArraysTotalElementCount","ArraysTotalSize"}]&;actionClass={"Class",actions(*{Left,Right,Up,Down}*)};loss=NetGraph[{1CrossEntropyLossLayer["Index"],2ThreadingLayer[#1*#2&]},{NetPort["DiscountedReward"]2,{NetPort["Input"],NetPort["Action"]}12},"Action"NetEncoder[actionClass]]loss//netSize

Out[]=

NetGraph