## Proximal Policy Optimization (PPO) Agents

Proximal policy optimization (PPO) is a model-free, online, on-policy, policy gradient reinforcement learning method. This algorithm is a type of policy gradient training that alternates between sampling data through environmental interaction and optimizing a clipped surrogate objective function using stochastic gradient descent. The clipped surrogate objective function improves training stability by limiting the size of the policy change at each step [1].

PPO is a simplified version of TRPO. TRPO is more computationally expensive than PPO, but TRPO tends to be more robust than PPO if the environment dynamics are deterministic and the observation is low dimensional. For more information on TRPO agents, see Trust Region Policy Optimization (TRPO) Agents.

For more information on the different types of reinforcement learning agents, see Reinforcement Learning Agents.

PPO agents can be trained in environments with the following observation and action spaces.

Observation Space | Action Space |
---|---|

Discrete or continuous | Discrete or continuous |

PPO agents use the following actor and critics.

Critic | Actor |
---|---|

Value function critic | Stochastic policy actor |

During training, a PPO agent:

Estimates probabilities of taking each action in the action space and randomly selects actions based on the probability distribution.

Interacts with the environment for multiple steps using the current policy before using mini-batches to update the actor and critic properties over multiple epochs.

If the `UseExplorationPolicy`

option of the agent is set to
`false`

the action with maximum likelihood is always used in `sim`

and `generatePolicyFunction`

. As a result, the simulated agent and generated policy
behave deterministically.

If the `UseExplorationPolicy`

is set to `true`

the
agent selects its actions by sampling its probability distribution. As a result the policy is
stochastic and the agent explores its observation space.

This option affects only simulation and deployment; it does not affect training.

### Actor and Critic Function Approximators

To estimate the policy and value function, a PPO agent maintains two function approximators.

Actor

*π*(*A*|*S*;*θ*) — The actor, with parameters*θ*, outputs the conditional probability of taking each action*A*when in state*S*as one of the following:Discrete action space — The probability of taking each discrete action. The sum of these probabilities across all actions is 1.

Continuous action space — The mean and standard deviation of the Gaussian probability distribution for each continuous action.

Critic

*V*(*S*;*ϕ*) — The critic, with parameters*ϕ*, takes observation*S*and returns the corresponding expectation of the discounted long-term reward.

For more information on creating actors and critics for function approximation, see Create Policies and Value Functions.

During training, the agent tunes the parameter values in *θ*. After
training, the parameters remain at their tuned value and the trained actor function
approximator is stored in
*π*(*A*|*S*).

### Agent Creation

You can create and train PPO agents at the MATLAB^{®} command line or using the **Reinforcement Learning Designer** app. For
more information on creating agents using **Reinforcement Learning Designer**, see
Create Agents Using Reinforcement Learning Designer.

At the command line, you can create a PPO agent with default actor and critic based on the observation and action specifications from the environment. To do so, perform the following steps.

Create observation specifications for your environment. If you already have an environment interface object, you can obtain these specifications using

`getObservationInfo`

.Create action specifications for your environment. If you already have an environment interface object, you can obtain these specifications using

`getActionInfo`

.If needed, specify the number of neurons in each learnable layer or whether to use an LSTM layer. To do so, create an agent initialization option object using

`rlAgentInitializationOptions`

.Specify agent options using an

`rlPPOAgentOptions`

object.Create the agent using an

`rlPPOAgent`

object.

Alternatively, you can create actor and critic and use these objects to create your agent. In this case, ensure that the input and output dimensions of the actor and critic match the corresponding action and observation specifications of the environment.

Create an actor using an

`rlDiscreteCategoricalActor`

object (for discrete action spaces) or an`rlContinuousGaussianActor`

object (for continuous action spaces).Create a critic using an

`rlValueFunction`

object.If needed, specify agent options using an

`rlPPOAgentOptions`

object.Create the agent using the

`rlPPOAgent`

function.

PPO agents support actors and critics that use recurrent deep neural networks as function approximators.

For more information on creating actors and critics for function approximation, see Create Policies and Value Functions.

### Training Algorithm

PPO agents use the following training algorithm. To configure the training algorithm,
specify options using an `rlPPOAgentOptions`

object.

Initialize the actor

*π*(*A*|*S*;*θ*) with random parameter values*θ*.Initialize the critic

*V*(*S*;*ϕ*) with random parameter values*ϕ*.Generate

*N*experiences by following the current policy. The experience sequence is$${S}_{ts},{A}_{ts},{R}_{ts+1},{S}_{ts+1},\dots ,{S}_{ts+N-1},{A}_{ts+N-1},{R}_{ts+N},{S}_{ts+N}$$

Here,

*S*is a state observation,_{t}*A*is an action taken from that state,_{t}*S*is the next state, and_{t+1}*R*is the reward received for moving from_{t+1}*S*to_{t}*S*._{t+1}When in state

*S*, the agent computes the probability of taking each action in the action space using_{t}*π*(*A*|*S*;_{t}*θ*) and randomly selects action*A*based on the probability distribution._{t}*ts*is the starting time step of the current set of*N*experiences. At the beginning of the training episode,*ts*= 1. For each subsequent set of*N*experiences in the same training episode,*ts*←*ts*+*N*.For each experience sequence that does not contain a terminal state,

*N*is equal to the`ExperienceHorizon`

option value. Otherwise,*N*is less than`ExperienceHorizon`

and*S*is the terminal state._{N}For each episode step

*t*=*ts*+1,*ts*+2, …,*ts*+*N*, compute the return and advantage function using the method specified by the`AdvantageEstimateMethod`

option.**Finite Horizon**(`AdvantageEstimateMethod = "finite-horizon"`

) — Compute the return*G*, which is the sum of the reward for that step and the discounted future reward [2]._{t}$${G}_{t}={\displaystyle \sum _{k=t}^{ts+N}\left({\gamma}^{k-t-1}{R}_{k}\right)}+b{\gamma}^{ts+N-t}V\left({S}_{ts+N};\varphi \right)$$

Here,

*b*is`0`

if*S*is a terminal state and_{ts+N}`1`

otherwise. That is, if*S*is not a terminal state, the discounted future reward includes the discounted state value function, computed using the critic network_{ts+N}*V*.Compute the advantage function

*D*._{t}$${D}_{t}={G}_{t}-V\left({S}_{t};\varphi \right)$$

**Generalized Advantage Estimator**(`AdvantageEstimateMethod = "gae"`

) — Compute the advantage function*D*, which is the discounted sum of temporal difference errors [3]._{t}$$\begin{array}{c}{D}_{t}={\displaystyle \sum _{k=t}^{ts+N-1}{\left(\gamma \lambda \right)}^{k-t}{\delta}_{k}}\\ {\delta}_{k}={R}_{t}+b\gamma V\left({S}_{t};\varphi \right)\end{array}$$

Here,

*b*is`0`

if*S*is a terminal state and_{ts+N}`1`

otherwise.*λ*is a smoothing factor specified using the`GAEFactor`

option.Compute the return

*G*._{t}$${G}_{t}={D}_{t}+V\left({S}_{t};\varphi \right)$$

To specify the discount factor

*γ*for either method, use the`DiscountFactor`

option.Learn from mini-batches of experiences over

*K*epochs. To specify*K*, use the`NumEpoch`

option. For each learning epoch:Sample a random mini-batch data set of size

*M*from the current set of experiences. To specify*M*, use the`MiniBatchSize`

option. Each element of the mini-batch data set contains a current experience and the corresponding return and advantage function values.Update the critic parameters by minimizing the loss

*L*across all sampled mini-batch data._{critic}$${L}_{critic}\left(\varphi \right)=\frac{1}{2M}{\displaystyle \sum _{i=1}^{M}{\left({G}_{i}-V\left({S}_{i};\varphi \right)\right)}^{2}}$$

Normalize the advantage values

*D*based on recent unnormalized advantage values._{i}If the

`NormalizedAdvantageMethod`

option is`'none'`

, do not normalize the advantage values.$${\widehat{D}}_{i}\leftarrow {D}_{i}$$

If the

`NormalizedAdvantageMethod`

option is`'current'`

, normalize the advantage values based on the unnormalized advantages in the current mini-batch.$${\widehat{D}}_{i}\leftarrow \frac{{D}_{i}-mean\left({D}_{1},{D}_{2},\dots ,{D}_{M}\right)}{std\left({D}_{1},{D}_{2},\dots ,{D}_{M}\right)}$$

If the

`NormalizedAdvantageMethod`

option is`'moving'`

, normalize the advantage values based on the unnormalized advantages for the*N*most recent advantages, including the current advantage value. To specify the window size*N*, use the`AdvantageNormalizingWindow`

option.$${\widehat{D}}_{i}\leftarrow \frac{{D}_{i}-mean\left({D}_{1},{D}_{2},\dots ,{D}_{N}\right)}{std\left({D}_{1},{D}_{2},\dots ,{D}_{N}\right)}$$

Update the actor parameters by minimizing the actor loss function

*L*across all sampled mini-batch data._{actor}$$\begin{array}{c}{L}_{actor}\left(\theta \right)=\frac{1}{M}{\displaystyle \sum _{i=1}^{M}\left(-\mathrm{min}\left({r}_{i}\left(\theta \right)\cdot {D}_{i},{c}_{i}\left(\theta \right)\cdot {D}_{i}\right)+w{\mathscr{H}}_{i}\left(\theta ,{S}_{i}\right)\right)}\\ {r}_{i}\left(\theta \right)=\frac{\pi \left({A}_{i}|{S}_{i};\theta \right)}{\pi \left({A}_{i}|{S}_{i};{\theta}_{old}\right)}\\ {c}_{i}\left(\theta \right)=\mathrm{max}\left(\mathrm{min}\left({r}_{i}\left(\theta \right),1+\epsilon \right),1-\epsilon \right)\end{array}$$

Here:

*D*and_{i}*G*are the advantage function and return value for the_{i}*i*th element of the mini-batch, respectively.*π*(*A*|_{i}*S*;_{i}*θ*) is the probability of taking action*A*when in state_{i}*S*, given the updated policy parameters_{i}*θ*.*π*(*A*|_{i}*S*;_{i}*θ*) is the probability of taking action_{old}*A*when in state_{i}*S*, given the previous policy parameters_{i}*θ*from before the current learning epoch._{old}*ε*is the clip factor specified using the`ClipFactor`

option.*ℋ*(_{i}*θ*) is the entropy loss and*w*is the entropy loss weight factor, specified using the`EntropyLossWeight`

option. For more information on entropy loss, see Entropy Loss.

Repeat steps 3 through 5 until the training episode reaches a terminal state.

### Entropy Loss

To promote agent exploration, you can add an entropy loss term
*w**ℋ _{i}*(

*θ*,

*S*) to the actor loss function, where

_{i}*w*is the entropy loss weight and

*ℋ*(

_{i}*θ*,

*S*) is the entropy.

_{i}The entropy value is higher when the agent is more uncertain about which action to take next. Therefore, maximizing the entropy loss term (minimizing the negative entropy loss) increases the agent uncertainty, thus encouraging exploration. To promote additional exploration, which can help the agent move out of local optima, you can specify a larger entropy loss weight.

For a discrete action space, the agent uses the following entropy value. In this case, the actor outputs the probability of taking each possible discrete action.

$${\mathscr{H}}_{i}\left(\theta ,{S}_{i}\right)=-{\displaystyle \sum _{k=1}^{P}\pi \left({A}_{k}|{S}_{i};\theta \right)\mathrm{ln}\pi \left({A}_{k}|{S}_{i};\theta \right)}$$

Here:

*P*is the number of possible discrete actions.*π*(*A*|_{k}*S*;_{i}*θ*) is the probability of taking action*A*when in state_{k}*S*following the current policy._{i}

For a continuous action space, the agent uses the following entropy value. In this case, the actor outputs the mean and standard deviation of the Gaussian distribution for each continuous action.

$${\mathscr{H}}_{i}\left(\theta ,{S}_{i}\right)=\frac{1}{2}{\displaystyle \sum _{k=1}^{C}\mathrm{ln}\left(2\pi \cdot e\cdot {\sigma}_{k,i}^{2}\right)}$$

Here:

*C*is the number of continuous actions output by the actor.*σ*is the standard deviation for action_{k,i}*k*when in state*S*following the current policy._{i}

## References

[1] Schulman, John, Filip Wolski,
Prafulla Dhariwal, Alec Radford, and Oleg Klimov. “Proximal Policy Optimization Algorithms.”
*ArXiv:1707.06347 [Cs]*, July 19, 2017. https://arxiv.org/abs/1707.06347.

[2] 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]*, February 4, 2016. https://arxiv.org/abs/1602.01783.

[3] Schulman, John, Philipp Moritz,
Sergey Levine, Michael Jordan, and Pieter Abbeel. “High-Dimensional Continuous Control Using
Generalized Advantage Estimation.” *ArXiv:1506.02438 [Cs]*, October 20,
2018. https://arxiv.org/abs/1506.02438.