Main Content

Deploy Trained Reinforcement Learning Policies

Once you train a reinforcement learning agent, you can generate code to deploy the optimal policy. You can generate:

  • CUDA® code for deep neural network policies using GPU Coder™

  • C/C++ code for table, deep neural network, or linear basis function policies using MATLAB® Coder™

Code generation is supported for agents using feedforward neural networks in any of the input paths, provided that all the used layers are supported. Code generation is not supported for continuous actions PG, AC, PPO, and SAC agents using a recurrent neural network (RNN).

For more information on training reinforcement learning agents, see Train Reinforcement Learning Agents.

To create a policy evaluation function that selects an action based on a given observation, use the generatePolicyFunction command. This command generates a MATLAB script, which contains the policy evaluation function, and a MAT-file, which contains the optimal policy data.

You can generate code to deploy this policy function using GPU Coder or MATLAB Coder.

Generate Code Using GPU Coder

If your trained optimal policy uses a deep neural network, you can generate CUDA code for the policy using GPU Coder. For more information on supported GPUs see GPU Support by Release (Parallel Computing Toolbox). There are several required and recommended prerequisite products for generating CUDA code for deep neural networks. For more information, see Installing Prerequisite Products (GPU Coder) and Setting Up the Prerequisite Products (GPU Coder).

Not all deep neural network layers support GPU code generation. For a list of supported layers, see Supported Networks, Layers, and Classes (GPU Coder). For more information and examples on GPU code generation, see Deep Learning with GPU Coder (GPU Coder).

Generate CUDA Code for Deep Neural Network Policy

As an example, generate GPU code for the policy gradient agent trained in Train PG Agent to Balance Cart-Pole System.

Load the trained agent.

load('MATLABCartpolePG.mat','agent')

Create a policy evaluation function for this agent.

generatePolicyFunction(agent)

This command creates the evaluatePolicy.m file, which contains the policy function, and the agentData.mat file, which contains the trained deep neural network actor. For a given observation, the policy function evaluates a probability for each potential action using the actor network. Then, the policy function randomly selects an action based on these probabilities.

You can generate code for this network using GPU Coder. For example, you can generate a CUDA compatible MEX function.

Configure the codegen function to create a CUDA compatible C++ MEX function.

cfg = coder.gpuConfig('mex');
cfg.TargetLang = 'C++';
cfg.DeepLearningConfig = coder.DeepLearningConfig('cudnn');

Set an example input value for the policy evaluation function. To find the observation dimension, use the getObservationInfo function. In this case, the observations are in a four-element vector.

argstr = '{ones(4,1)}';

Generate code using the codegen function.

codegen('-config','cfg','evaluatePolicy','-args',argstr,'-report');

This command generates the MEX function evaluatePolicy_mex.

Generate Code Using MATLAB Coder

You can generate C/C++ code for table, deep neural network, or linear basis function policies using MATLAB Coder.

Using MATLAB Coder, you can generate:

Generate C Code for Deep Neural Network Policy without using any Third-Party Library

As an example, generate C code without dependencies on third-party libraries for the policy gradient agent trained in Train PG Agent to Balance Cart-Pole System.

Load the trained agent.

load('MATLABCartpolePG.mat','agent')

Create a policy evaluation function for this agent.

generatePolicyFunction(agent)

This command creates the evaluatePolicy.m file, which contains the policy function, and the agentData.mat file, which contains the trained deep neural network actor. For a given observation, the policy function evaluates a probability for each potential action using the actor network. Then, the policy function randomly selects an action based on these probabilities.

Configure the codegen function to generate code suitable for building a MEX file.

cfg = coder.config('mex');

On the configuration object, set the target language to C++, and set DeepLearningConfig to 'none'. This option generates code without using any third-party library.

cfg.TargetLang = 'C';
cfg.DeepLearningConfig = coder.DeepLearningConfig('none'); 

Set an example input value for the policy evaluation function. To find the observation dimension, use the getObservationInfo function. In this case, the observations are in a four-element vector.

argstr = '{ones(4,1)}';

Generate code using the codegen function.

codegen('-config','cfg','evaluatePolicy','-args',argstr,'-report');

This command generates the C++ code for the policy gradient agent containing the deep neural network actor.

Generate C++ Code for Deep Neural Network Policy using Third-Party Libraries

As an example, generate C++ code for the policy gradient agent trained in Train PG Agent to Balance Cart-Pole System using the Intel Math Kernel Library for Deep Neural Networks (MKL-DNN).

Load the trained agent.

load('MATLABCartpolePG.mat','agent')

Create a policy evaluation function for this agent.

generatePolicyFunction(agent)

This command creates the evaluatePolicy.m file, which contains the policy function, and the agentData.mat file, which contains the trained deep neural network actor. For a given observation, the policy function evaluates a probability for each potential action using the actor network. Then, the policy function randomly selects an action based on these probabilities.

Configure the codegen function to generate code suitable for building a MEX file.

cfg = coder.config('mex');

On the configuration object, set the target language to C++, and set DeepLearningConfig to the target library 'mkldnn'. This option generates code using the Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN).

cfg.TargetLang = 'C++';
cfg.DeepLearningConfig = coder.DeepLearningConfig('mkldnn'); 

Set an example input value for the policy evaluation function. To find the observation dimension, use the getObservationInfo function. In this case, the observations are in a four-element vector.

argstr = '{ones(4,1)}';

Generate code using the codegen function.

codegen('-config','cfg','evaluatePolicy','-args',argstr,'-report');

This command generates the C++ code for the policy gradient agent containing the deep neural network actor.

Generate C Code for Q Table Policy

As an example, generate C code for the Q-learning agent trained in Train Reinforcement Learning Agent in Basic Grid World.

Load the trained agent.

load('basicGWQAgent.mat','qAgent')

Create a policy evaluation function for this agent.

generatePolicyFunction(qAgent)

This command creates the evaluatePolicy.m file, which contains the policy function, and the agentData.mat file, which contains the trained Q table value function. For a given observation, the policy function looks up the value function for each potential action using the Q table. Then, the policy function selects the action for which the value function is greatest.

Set an example input value for the policy evaluation function. To find the observation dimension, use the getObservationInfo function. In this case, there is a single one dimensional observation (belonging to a discrete set of possible values).

argstr = '{[1]}';

Configure the codegen function to generate embeddable C code suitable for targeting a static library, and set the output folder to buildFolder.

cfg = coder.config('lib');
outFolder = 'buildFolder';

Generate C code using the codegen function.

codegen('-c','-d',outFolder,'-config','cfg',...
    'evaluatePolicy','-args',argstr,'-report');

See Also

Related Topics