main content

evaluate function approximator object given observation (or observation-凯发k8网页登录

evaluate function approximator object given observation (or observation-action) input data

since r2022a

syntax

outdata = evaluate(fcnappx,indata)
[outdata,state] = evaluate(fcnappx,indata)

description

outdata = evaluate(fcnappx,indata) evaluates the function approximator object (that is, the actor or critic) fcnappx given the input value indata. it returns the output value outdata.

example

[outdata,state] = evaluate(fcnappx,indata) also returns the updated state of fcnappx when it contains a recurrent neural network.

examples

this example shows you how to evaluate a function approximator object (that is, an actor or a critic). for this example, the function approximator object is a discrete categorical actor and you evaluate it given some observation data, obtaining in return the action probability distribution and the updated network state.

load the same environment used in , and obtain the observation and action specifications.

env = rlpredefinedenv("cartpole-discrete");
obsinfo = getobservationinfo(env)
obsinfo = 
  rlnumericspec with properties:
     lowerlimit: -inf
     upperlimit: inf
           name: "cartpole states"
    description: "x, dx, theta, dtheta"
      dimension: [4 1]
       datatype: "double"

actinfo = getactioninfo(env)
actinfo = 
  rlfinitesetspec with properties:
       elements: [-10 10]
           name: "cartpole action"
    description: [0x0 string]
      dimension: [1 1]
       datatype: "double"

to approximate the policy within the actor, use a recurrent deep neural network. define the network as an array of layer objects. get the dimensions of the observation space and the number of possible actions directly from the environment specification objects. 

net = [
    sequenceinputlayer(prod(obsinfo.dimension))
    fullyconnectedlayer(8)
    relulayer
    lstmlayer(8,outputmode="sequence")
    fullyconnectedlayer(numel(actinfo.elements)) ];

convert the network to a dlnetwork object and display the number of weights.

net = dlnetwork(net);
summary(net)
   initialized: true
   number of learnables: 602
   inputs:
      1   'sequenceinput'   sequence input with 4 dimensions

create a stochastic actor representation for the network.

actor = rldiscretecategoricalactor(net,obsinfo,actinfo);

use evaluate to return the probability of each of the two possible actions. note that the type of the returned numbers is single, not double.

[prob,state] = evaluate(actor,{rand(obsinfo.dimension)});
prob{1}
ans = 2x1 single column vector
    0.4847
    0.5153

since a recurrent neural network is used for the actor, the second output argument, representing the updated state of the neural network, is not empty. in this case, it contains the updated (cell and hidden) states for the eight units of the lstm layer used in the network.

state{:}
ans = 8x1 single column vector
   -0.0833
    0.0619
   -0.0066
   -0.0651
    0.0714
   -0.0957
    0.0614
   -0.0326

ans = 8x1 single column vector
   -0.1367
    0.1142
   -0.0158
   -0.1820
    0.1305
   -0.1779
    0.0947
   -0.0833

you can use getstate and setstate to extract and set the current state of the actor.

getstate(actor)
ans=2×1 cell array
    {8x1 single}
    {8x1 single}

actor = setstate(actor, ...
                    {-0.01*single(rand(8,1)), ...
                      0.01*single(rand(8,1))});

you can obtain action probabilities and updated states for a batch of observations. for example, use a batch of five independent observations.

obsbatch = reshape(1:20,4,1,5,1);
[prob,state] = evaluate(actor,{obsbatch})
prob = 1x1 cell array
    {2x5 single}

state=2×1 cell array
    {8x5 single}
    {8x5 single}

the output arguments contain action probabilities and updated states for each observation in the batch.

note that the actor treats observation data along the batch length dimension independently, not sequentially.

prob{1}
ans = 2x5 single matrix
    0.5187    0.5869    0.6048    0.6124    0.6155
    0.4813    0.4131    0.3952    0.3876    0.3845

prob = evaluate(actor,{obsbatch(:,:,[5 4 3 1 2])});
prob{1}
ans = 2x5 single matrix
    0.6155    0.6124    0.6048    0.5187    0.5869
    0.3845    0.3876    0.3952    0.4813    0.4131

to evaluate the actor using sequential observations, use the sequence length (time) dimension. for example, obtain action probabilities for five independent sequences, each one made of nine sequential observations.

[prob,state] = evaluate(actor, ...
    {rand([obsinfo.dimension 5 9])})
prob = 1x1 cell array
    {2x5x9 single}

state=2×1 cell array
    {8x5 single}
    {8x5 single}

the first output argument contains a vector of two probabilities (first dimension) for each element of the observation batch (second dimension) and for each time element of the sequence length (third dimension).

the second output argument contains two vectors of final states for each observation batch (that is, the network maintains a separate state history for each observation batch).

display the probability of the second action, after the seventh sequential observation in the fourth independent batch. 

prob{1}(2,4,7)
ans = single
    0.5675

for more information on input and output format for recurrent neural networks, see the algorithms section of .

input arguments

function approximator object, specified as one of the following:

input data for the function approximator, specified as a cell array with as many elements as the number of input channels of fcnappx. in the following section, the number of observation channels is indicated by no.

  • if fcnappx is an rlqvaluefunction, an rlcontinuousdeterministictransitionfunction or an rlcontinuousgaussiantransitionfunction object, then each of the first no elements of indata must be a matrix representing the current observation from the corresponding observation channel. they must be followed by a final matrix representing the action.

  • if fcnappx is a function approximator object representing an actor or critic (but not an rlqvaluefunction object), indata must contain no elements, each one a matrix representing the current observation from the corresponding observation channel.

  • if fcnappx is an rlcontinuousdeterministicrewardfunction, an rlcontinuousgaussianrewardfunction, or an rlisdonefunction object, then each of the first no elements of indata must be a matrix representing the current observation from the corresponding observation channel. they must be followed by a matrix representing the action, and finally by no elements, each one being a matrix representing the next observation from the corresponding observation channel.

each element of indata must be a matrix of dimension mc-by-lb-by-ls, where:

  • mc corresponds to the dimensions of the associated input channel.

  • lb is the batch size. to specify a single observation, set lb = 1. to specify a batch of (independent) inputs, specify lb > 1. if indata has multiple elements, then lb must be the same for all elements of indata.

  • ls specifies the sequence length (length of the sequence of inputs along the time dimension) for recurrent neural network. if fcnappx does not use a recurrent neural network (which is the case for environment function approximators, as they do not support recurrent neural networks), then ls = 1. if indata has multiple elements, then ls must be the same for all elements of indata.

for more information on input and output formats for recurrent neural networks, see the algorithms section of .

example: {rand(8,3,64,1),rand(4,1,64,1),rand(2,1,64,1)}

output arguments

output data from the evaluation of the function approximator object, returned as a cell array. the size and contents of outdata depend on the type of object you use for fcnappx, and are shown in the following list. here, no is the number of observation channels.

  • rlcontinuousdeterministictransitionfunction - no matrices, each one representing the predicted observation from the corresponding observation channel.

  • rlcontinuousgaussiantransitionfunction - no matrices representing the mean value of the predicted observation for the corresponding observation channel, followed by no matrices representing the standard deviation of the predicted observation for the corresponding observation channel.

  • rlcontinuousgaussianactor - two matrices representing the mean value and standard deviation of the action, respectively.

  • rldiscretecategoricalactor - a matrix with the probabilities of each action.

  • rlcontinuousdeterministicactor a matrix with the action.

  • rlvectorqvaluefunction - a matrix with the values of each possible action.

  • rlqvaluefunction - a matrix with the value of the action.

  • rlvaluefunction - a matrix with the value of the current observation.

  • rlcontinuousdeterministicrewardfunction - a matrix with the predicted reward as a function of current observation, action, and next observation following the action.

  • rlcontinuousgaussianrewardfunction - two matrices representing the mean value and standard deviation, respectively, of the predicted reward as a function of current observation, action, and next observation following the action.

  • rlisdonefunction - a vector with the probabilities of the predicted termination status. termination probabilities range from 0 (no termination predicted) or 1 (termination predicted), and depend (in the most general case) on the values of observation, action, and next observation following the action.

each element of outdata is a matrix of dimensions d-by-lb-by-ls, where:

  • d is the vector of dimensions of the corresponding output channel of fcnappx. depending on the type of approximator function, this channel can carry a predicted observation (or its mean value or standard deviation), an action (or its mean value or standard deviation), the value (or values) of an observation (or observation-action couple), a predicted reward, or a predicted termination status.

  • lb is the batch size (length of a batch of independent inputs).

  • ls is the sequence length (length of the sequence of inputs along the time dimension) for a recurrent neural network. if fcnappx does not use a recurrent neural network (which is the case for environment function approximators, as they do not support recurrent neural networks), then ls = 1.

note

if fcnappx is a critic, then evaluate behaves identically to except that it returns results inside a single-cell array. if fcnappx is an rlcontinuousdeterministicactor actor, then evaluate behaves identically to . if fcnappx is a stochastic actor such as an or , then evaluate returns the action probability distribution, while returns a sample action. specifically, for an actor object, evaluate returns the probability of each possible action. for an actor object, evaluate returns the mean and standard deviation of the gaussian distribution. for these kinds of actors, see also the note in regarding the enforcement of constraints set by the action specification.

note

if fcnappx is an rlcontinuousdeterministicrewardfunction object, then evaluate behaves identically to predict except that it returns results inside a single-cell array. if fcnappx is an rlcontinuousdeterministictransitionfunction object, then evaluate behaves identically to predict. if fcnappx is an rlcontinuousgaussiantransitionfunction object, then evaluate returns the mean value and standard deviation the observation probability distribution, while predict returns an observation sampled from this distribution. similarly, for an rlcontinuousgaussianrewardfunction object, evaluate returns the mean value and standard deviation the reward probability distribution, while predict returns a reward sampled from this distribution. finally, if fcnappx is an rlisdonefunction object, then evaluate returns the probabilities of the termination status being false or true, respectively, while predict returns a predicted termination status sampled with these probabilities.

next state of the function approximator object, returned as a cell array. if fcnappx does not use a recurrent neural network (which is the case for environment function approximators), then state is an empty cell array.

you can set the state of the approximator to state using the setstate function. for example:

critic = setstate(critic,state);

version history

introduced in r2022a

see also

functions

objects

topics

网站地图