Options for DQN agent
Use an rlDQNAgentOptions object to specify options for deep
Q-network (DQN) agents. To create a DQN agent, use rlDQNAgent.
For more information, see Deep Q-Network Agents.
For more information on the different types of reinforcement learning agents, see Reinforcement Learning Agents.
opt = rlDQNAgentOptions
opt = rlDQNAgentOptions(Name,Value)rlDQNAgentOptions('DiscountFactor',0.95)
creates an option set with a discount factor of 0.95. You can specify
multiple name-value pairs. Enclose each property name in quotes.
UseDoubleDQN — Flag for using double DQNFlag for using double DQN for value function target updates, specified as a logical
value. For most application set UseDoubleDQN to
"on". For more information, see Deep Q-Network Agents.
EpsilonGreedyExploration — Options for epsilon-greedy explorationEpsilonGreedyExploration objectOptions for epsilon-greedy exploration, specified as an
EpsilonGreedyExploration object with the following
properties.
| Property | Description | Default Value |
|---|---|---|
Epsilon | Probability threshold to either randomly select an action or select the
action that maximizes the state-action value function. A larger value of
Epsilon means that the agent randomly explores the action
space at a higher rate. | 1 |
EpsilonMin | Minimum value of Epsilon | 0.01 |
EpsilonDecay | Decay rate | 0.0050 |
At the end of each training time step, if Epsilon is greater than
EpsilonMin, then it is updated using the following formula.
Epsilon = Epsilon*(1-EpsilonDecay)
To specify exploration options, use dot notation after creating the
rlDQNAgentOptions object. For example, set the epsilon value to
0.9.
opt = rlDQNAgentOptions; opt.EpsilonGreedyExploration.Epsilon = 0.9;
If your agent converges on local optima too quickly, promote agent exploration by
increasing Epsilon.
SequenceLength — Maximum batch-training trajectory length when using RNN1 (default) | positive integerMaximum batch-training trajectory length when using a recurrent neural network for
the critic, specified as a positive integer. This value must be greater than
1 when using a recurrent neural network for the critic and
1 otherwise.
TargetSmoothFactor — Smoothing factor for target critic updates1e-3 (default) | positive scalar less than or equal to 1Smoothing factor for target critic updates, specified as a positive scalar less than or equal to 1. For more information, see Target Update Methods.
TargetUpdateFrequency — Number of steps between target critic updates1 (default) | positive integerNumber of steps between target critic updates, specified as a positive integer. For more information, see Target Update Methods.
ResetExperienceBufferBeforeTraining — Flag for clearing the experience buffertrue (default) | falseFlag for clearing the experience buffer before training, specified as a logical value.
SaveExperienceBufferWithAgent — Flag for saving the experience bufferfalse (default) | trueFlag for saving the experience buffer data when saving the agent, specified as a
logical value. This option applies both when saving candidate agents during training and
when saving agents using the save function.
For some agents, such as those with a large experience buffer and image-based
observations, the memory required for saving their experience buffer is large. In such
cases, to not save the experience buffer data, set
SaveExperienceBufferWithAgent to false.
If you plan to further train your saved agent, you can start training with the
previous experience buffer as a starting point. In this case, set
SaveExperienceBufferWithAgent to true.
MiniBatchSize — Size of random experience mini-batch64 (default) | positive integerSize of random experience mini-batch, specified as a positive integer. During each training episode, the agent randomly samples experiences from the experience buffer when computing gradients for updating the critic properties. Large mini-batches reduce the variance when computing gradients but increase the computational effort.
When using a recurrent neural network for the critic,
MiniBatchSize is the number of experience trajectories in a
batch, where each trajectory has length equal to
SequenceLength.
NumStepsToLookAhead — Number of steps ahead1 (default) | positive integerNumber of steps to look ahead during training, specified as a positive integer.
N-step Q learning is not supported when using a recurrent neural network for the
critic. In this case, NumStepsToLookAhead must be
1.
ExperienceBufferLength — Experience buffer size10000 (default) | positive integerExperience buffer size, specified as a positive integer. During training, the agent updates the critic using a mini-batch of experiences randomly sampled from the buffer.
SampleTime — Sample time of agent1 (default) | positive scalarSample time of agent, specified as a positive scalar.
Within a Simulink environment, the agent gets executed every
SampleTime seconds of simulation time.
Within a MATLAB environment, the agent gets executed every time the environment
advances. However, SampleTime is the time interval between
consecutive elements in the output experience returned by sim or
train.
DiscountFactor — Discount factor0.99 (default) | positive scalar less than or equal to 1Discount factor applied to future rewards during training, specified as a positive scalar less than or equal to 1.
rlDQNAgent | Deep Q-network reinforcement learning agent |
This example shows how to create a DQN agent options object.
Create an rlDQNAgentOptions object that specifies the agent mini-batch size.
opt = rlDQNAgentOptions('MiniBatchSize',48)opt =
rlDQNAgentOptions with properties:
UseDoubleDQN: 1
EpsilonGreedyExploration: [1x1 rl.option.EpsilonGreedyExploration]
SequenceLength: 1
TargetSmoothFactor: 1.0000e-03
TargetUpdateFrequency: 1
ResetExperienceBufferBeforeTraining: 1
SaveExperienceBufferWithAgent: 0
MiniBatchSize: 48
NumStepsToLookAhead: 1
ExperienceBufferLength: 10000
SampleTime: 1
DiscountFactor: 0.9900
You can modify options using dot notation. For example, set the agent sample time to 0.5.
opt.SampleTime = 0.5;
Behavior changed in R2020a
Target update method settings for DQN agents have changed. The following changes require updates to your code:
The TargetUpdateMethod option has been removed. Now, DQN agents
determine the target update method based on the TargetUpdateFrequency
and TargetSmoothFactor option values.
The default value of TargetUpdateFrequency has changed from
4 to 1.
To use one of the following target update methods, set the
TargetUpdateFrequency and TargetSmoothFactor
properties as indicated.
| Update Method | TargetUpdateFrequency | TargetSmoothFactor |
|---|---|---|
| Smoothing | 1 | Less than 1 |
| Periodic | Greater than 1 | 1 |
| Periodic smoothing (new method in R2020a) | Greater than 1 | Less than 1 |
The default target update configuration, which is a smoothing update with a
TargetSmoothFactor value of 0.001, remains the
same.
This table shows some typical uses of rlDQNAgentOptions
and how to update your code to use the new option configuration.
| Not Recommended | Recommended |
|---|---|
opt = rlDQNAgentOptions('TargetUpdateMethod',"smoothing"); |
opt = rlDQNAgentOptions; |
opt = rlDQNAgentOptions('TargetUpdateMethod',"periodic"); |
opt = rlDQNAgentOptions; opt.TargetUpdateFrequency = 4; opt.TargetSmoothFactor = 1; |
opt = rlDQNAgentOptions; opt.TargetUpdateMethod = "periodic"; opt.TargetUpdateFrequency = 5; |
opt = rlDQNAgentOptions; opt.TargetUpdateFrequency = 5; opt.TargetSmoothFactor = 1; |
You have a modified version of this example. Do you want to open this example with your edits?