train ac agent to balance cart-凯发k8网页登录

actor-critic agent parallel training

ac agents use gradient-based parallelization, in which both environment simulation and gradient computation is done by the workers and the gradient averaging is done by the client. specifically, each worker generates new experiences from its copy of agent and environment, calculates the gradient, and sends it back to the client. the client updates its parameters as follows.

for more information on gradient-based parallelization, see train agents using parallel computing and gpus.

create cart-pole matlab environment interface

create a predefined environment interface for the cart-pole system. for more information on this environment, see load predefined control system environments.

env = rlpredefinedenv("cartpole-discrete");
env.penaltyforfalling = -10;

obtain the observation and action information from the environment interface.

obsinfo = getobservationinfo(env);
numobservations = obsinfo.dimension(1);
actinfo = getactioninfo(env);

fix the random generator seed for reproducibility.

rng(0)

create ac agent

actor-critic agents use a parametrized value function approximator to estimate the value of the policy. a value-function critic takes the current observation as input and returns a single scalar as output (the estimated discounted cumulative long-term reward for following the policy from the state corresponding to the current observation).

to model the parametrized value function within the critic, use a neural network with one input layer (which receives the content of the observation channel, as specified by obsinfo) and one output layer (which returns the scalar value). note that prod(obsinfo.dimension) returns the total number of dimensions of the observation space regardless of whether the observation space is a column vector, row vector, or matrix.

define the network as an array of layer objects.

criticnetwork = [
    featureinputlayer(prod(obsinfo.dimension))
    fullyconnectedlayer(32)
    relulayer
    fullyconnectedlayer(1)
    ];

create the value function approximator object using criticnetwork and the environment action and observation specifications.

critic = rlvaluefunction(criticnetwork,obsinfo);

actor-critic agents use a parametrized stochastic policy, which for discrete action spaces is implemented by a discrete categorical actor. this actor takes an observation as input and returns as output a random action sampled (among the finite number of possible actions) from a categorical probability distribution.

to model the parametrized policy within the actor, use a neural network with one input layer (which receives the content of the environment observation channel, as specified by obsinfo) and one output layer. the output layer must return a vector of probabilities for each possible action, as specified by actinfo. note that numel(actinfo.dimension) returns the number of elements of the discrete action space.

define the network as an array of layer objects.

actornetwork = [
    featureinputlayer(prod(obsinfo.dimension))
    fullyconnectedlayer(32)
    relulayer
    fullyconnectedlayer(numel(actinfo.dimension))];

create the actor approximator object using actornetwork and the environment action and observation specifications.

actor = rldiscretecategoricalactor(actornetwork,obsinfo,actinfo);

for more information on creating approximator objects such as actors and critics, see create policies and value functions.

specify options for the critic and actor using rloptimizeroptions.

criticopts = rloptimizeroptions(learnrate=1e-2,gradientthreshold=1);
actoropts = rloptimizeroptions(learnrate=1e-2,gradientthreshold=1);

specify the ac agent options using rlacagentoptions, include the training options for the actor and critic.

agentopts = rlacagentoptions(...
    actoroptimizeroptions=actoropts,...
    criticoptimizeroptions=criticopts,...
    entropylossweight=0.01);

create the agent using the specified actor representation and agent options. for more information, see rlacagent.

agent = rlacagent(actor,critic,agentopts);

parallel training options

to train the agent, first specify the training options. for this example, use the following options.

trainopts = rltrainingoptions(...
    maxepisodes=1000,...
    maxstepsperepisode=500,...
    verbose=false,...
    plots="training-progress",...
    stoptrainingcriteria="averagereward",...
    stoptrainingvalue=500,...
    scoreaveragingwindowlength=10); 

you can visualize the cart-pole system can during training or simulation using the plot function.

plot(env)

to train the agent using parallel computing, specify the following training options.

trainopts.useparallel = true;
trainopts.parallelizationoptions.mode = "async";

for more information, see rltrainingoptions.

train agent

train the agent using the train function. training the agent is a computationally intensive process that takes several minutes to complete. to save time while running this example, load a pretrained agent by setting dotraining to false. to train the agent yourself, set dotraining to true. due to randomness in the asynchronous parallel training, you can expect different training results from the following training plot. the plot shows the result of training with six workers.

dotraining = false;
if dotraining    
    % train the agent.
    trainingstats = train(agent,env,trainopts);
else
    % load the pretrained agent for the example.
    load("matlabcartpoleparac.mat","agent");
end

simulate ac agent

you can visualize the cart-pole system with the plot function during simulation.

plot(env)

to validate the performance of the trained agent, simulate it within the cart-pole environment. for more information on agent simulation, see rlsimulationoptions and sim.

simoptions = rlsimulationoptions(maxsteps=500);
experience = sim(env,agent,simoptions);

totalreward = sum(experience.reward)

totalreward = 500

references

[1] mnih, volodymyr, adrià puigdomènech badia, mehdi mirza, alex graves, timothy p. lillicrap, tim harley, david silver, and koray kavukcuoglu. ‘asynchronous methods for deep reinforcement learning’. arxiv:1602.01783 [cs], 16 june 2016. .