Train Multiple Agents for Area Coverage

This example demonstrates a multi-agent collaborative task in which you train three proximal policy optimization (PPO) agents to achieve full coverage of a grid-world environment. As shown in this example, if you define your environment behavior using MATLAB code, you can also incorporate it into a Simulink® environment using MATLAB Function blocks.

Create Environment

The environment in this example is a 12x12 grid world containing obstacles, with unexplored cells marked in white and obstacles marked in black. The environment contains three robots, represented by the red, green, and blue circles. Three proximal policy optimization agents with discrete action spaces control the robots. To learn more about PPO agents, see Proximal Policy Optimization (PPO) Agents.

The agents provide one of five possible movement actions (WAIT, UP, DOWN, LEFT, or RIGHT) to their respective robots. The robots decide whether an action is legal or illegal. For example, an action of moving LEFT when the robot is located next to the left boundary of the environment is deemed illegal. Similarly, actions for colliding against obstacles and other agents in the environment are illegal actions and draw penalties. The overall goal is to explore all cells within a fixed time frame.

At each time step, an agent observes the environment using a 12x12x4 image. Each of the four channels in the image captures the following information in order:

• Cells with obstacles.

• Current position of the robot (self) being controlled.

• Position of other (friend) robots.

• Cells that have been explored during the episode.

The following figure shows an example of what the agent controlling the green robot observes for a given time step.

For the grid world environment:

• The search area is a 12x12 grid with obstacles.

• The observation for each agent is a 12x12x4 image.

• The discrete action set contains five actions (WAIT=0, UP=1, DOWN=2, LEFT=3, RIGHT=4).

• The simulation terminates when the grid is fully explored or the maximum number of steps is reached.

At each time step, agents receive the following rewards:

• +2 for moving to a previously unexplored cell (white).

• -1 for an illegal action (attempt to move outside the boundary or collide against other robots and obstacles).

• -1 for an action that results in no motion (lazy penalty).

• -0.5 for moving to an explored cell.

• +100 if the grid is fully explored.

Define the locations of obstacles within the grid using a matrix of indices. The first column contains the row indices, and the second column contains the column indices.

obsMat = [ ...
    4 3; ...
    5 3; ...
    6 3; ...
    7 3; ...
    8 3; ...
    9 3; ...
    5 11; ... 
    6 11; ...
    7 11; ...
    8 11; ...
    5 12; ...
    6 12; ...
    7 12; ...
    8 12 ];

Initialize the robot positions.

sA0 = [2 2];
sB0 = [11 4];
sC0 = [3 12];
s0 = [sA0; sB0; sC0];

Configure the initial condition of the grid.

% initial grid
g0 = zeros(12,12);

% obstacle cells have the value 1.0
for idx = 1:size(obsMat,1)
    r = obsMat(idx,1);
    c = obsMat(idx,2);
    g0(r,c) = 1.0;
end

% robot cells have the values 0.25 (red), 0.50 (green), and 0.75 (blue)
g0(sA0(1),sA0(2)) = 0.25;
g0(sB0(1),sB0(2)) = 0.50;
g0(sC0(1),sC0(2)) = 0.75;

Specify the sample time, simulation time, and maximum number of steps per episode.

Ts = 0.1;
Tf = 100;
maxsteps = ceil(Tf/Ts);

Open the model. There are three RL Agent blocks and a subsystem block that models the area coverage environment.

mdl = "rlAreaCoverage";
open_system(mdl)

The environment dynamics is implemented using MATLAB Function blocks. Navigate to the Environment subsystem to view the implementation.

open_system(mdl+"/Environment")

Note that the states of the environment (grid, numSteps, numExplored and nextX signals) are updated iteratively in the feedback loop. Initial values for those signals are specified in the unit delay blocks to ensure the environment is correctly initialized.

In this example, the agents are homogeneous and have the same observation and action specifications. Create the observation and action specifications for the environment. For more information, see rlNumericSpec and rlFiniteSetSpec.

% Define observation specifications.
obsSize = [12 12 4];
oinfo = rlNumericSpec(obsSize);
oinfo.Name = "observations";

% Define action specifications.
numAct = 5;
actionSpace = {0,1,2,3,4};
ainfo = rlFiniteSetSpec(actionSpace);
ainfo.Name = "actions";

Specify the block paths for the agents.

blks = mdl + ["/Agent A (Red)","/Agent B (Green)","/Agent C (Blue)"];

Create the environment interface, specifying the same observation and action specifications for all three agents.

env = rlSimulinkEnv(mdl,blks,{oinfo,oinfo,oinfo},{ainfo,ainfo,ainfo});

Specify a reset function for the environment. The function resetMap (provided at the end of this example) ensures that the robots start from random initial positions at the beginning of each episode. The random initialization makes the agents robust to different starting positions and can improve training performance.

env.ResetFcn = @(in) resetMap(in, obsMat);

Create Agents

The PPO agents in this example operate on a discrete action space and rely on actor and critic functions to learn the optimal policies. The agents maintain deep neural network-based function approximators for the actors and critics with similar network structures (a combination of convolution and fully connected layers). The critic outputs a scalar value representing the state value $\mathit{V}\left(\mathit{s}\right)$. The actor outputs the probabilities $\pi \left(\mathit{a}|\mathit{s}\right)$ of taking each of the five actions WAIT, UP, DOWN, LEFT, or RIGHT.

Set the random seed for reproducibility.

rng(0)

Create the actor and critic functions using the following steps.

1. Create the actor and critic deep neural networks.

2. Create the actor function objects using the rlDiscreteCategoricalActor command.

3. Create the critic function objects using the rlValueFunction command.

Use the same network structure and representation options for all three agents.

% Create actor deep neural network.
actorNetWork = [
    imageInputLayer(obsSize)
    convolution2dLayer(3,16,Padding="same",PaddingValue=-1,Stride=2)
    reluLayer
    maxPooling2dLayer(3,Padding="same",Stride=2)
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(numAct)
    softmaxLayer
    ];
actorNetWork = dlnetwork(actorNetWork);

% Create critic deep neural network.
criticNetwork = [
    imageInputLayer(obsSize)
    convolution2dLayer(3,16,Padding="same",PaddingValue=-1,Stride=2)
    reluLayer
    maxPooling2dLayer(3,Padding="same",Stride=2)
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(1)
    ];
criticNetwork = dlnetwork(criticNetwork);

Create the actor and critic objects.

for idx = 1:3
    actor(idx) = rlDiscreteCategoricalActor(actorNetWork,oinfo,ainfo);
    critic(idx) = rlValueFunction(criticNetwork,oinfo);
end

Specify training options for the critic and the actor using rlOptimizerOptions. The agents learn using the learning rate 1e-3, gradient threshold 1.0, and L2 regularization factor 1e-4.

actorOpts  = rlOptimizerOptions(LearnRate=1e-3, GradientThreshold=1, L2RegularizationFactor=1e-4);
criticOpts = rlOptimizerOptions(LearnRate=1e-3, GradientThreshold=1, L2RegularizationFactor=1e-4);

Specify the agent options using rlPPOAgentOptions, include the training options for the actor and critic. Use the same options for all three agents.

opt = rlPPOAgentOptions(...
    ActorOptimizerOptions=actorOpts,...
    CriticOptimizerOptions=criticOpts,...
    ExperienceHorizon=500,...
    MiniBatchSize=200,...
    SampleTime=Ts,...
    DiscountFactor=0.9);

During training, the agents collect trajectories and train using the mini-batch size of 200. An experience horizon of 500 steps is used for advantage computation, and the discount factor value of 0.9 encourages long-term rewards.

Create the agents using the defined actors, critics, and agent options.

agentA = rlPPOAgent(actor(1),critic(1),opt);
agentB = rlPPOAgent(actor(2),critic(2),opt);
agentC = rlPPOAgent(actor(3),critic(3),opt);

Alternatively, you can create the agents first, and then access their option object to modify the options using dot notation.

Train Agents

In this example, the agents are trained in centralized manner. Specify the following options for training the agents.

• Specify all agents under one agent group.

• Specify the "centralized" learning strategy.

• Run the training for at most 500 episodes, with each episode lasting at most 1000 time steps.

• Stop the training an agent when its average reward over 20 consecutive episodes is 200 or more.

trainOpts = rlMultiAgentTrainingOptions(...
    "AgentGroups",{[1,2,3]},...
    "LearningStrategy","centralized",...
    MaxEpisodes=500,...
    MaxStepsPerEpisode=maxsteps,...
    SimulationStorageType="none",...
    ScoreAveragingWindowLength=20,...
    StopTrainingCriteria="AverageReward", ...
    StopTrainingValue=200);

For more information on multi-agent training, see rlMultiAgentTrainingOptions.

To train the agents, specify an array of agents to the train function. The order of the agents in the array must match the order of agent block paths specified during environment creation. Doing so ensures that the agent objects are linked to the appropriate action and observation specifications in the environment.

Training is a computationally intensive process that takes several minutes to complete. To save time while running this example, load pretrained agent parameters by setting doTraining to false. To train the agent yourself, set doTraining to true.

doTraining = false;
if doTraining
    result = train([agentA,agentB,agentC],env,trainOpts);
else
    load("rlAreaCoverageAgents.mat","params");
    % The saved agents were trained in centralized manner with synchronized
    % model parameters. Set the trained parameters on all agents.
    setLearnableParameters(agentA, params);
    setLearnableParameters(agentB, params);
    setLearnableParameters(agentC, params);
end

The following figure shows a snapshot of the training progress. You can expect different results due to randomness in the training process.

Simulate Agents

Reset the random seed for reproducibility.

rng(0)

Simulate the trained agents within the environment. For more information on agent simulation, see rlSimulationOptions and sim.

simOpts = rlSimulationOptions(MaxSteps=maxsteps);
experience = sim(env,[agentA,agentB,agentC],simOpts);

The agents are able to achieve full coverage of the grid world.

Local Functions

function in = resetMap(in,obsMat)
% Reset function for multi agent area coverage example

gridSize = [12 12];
isvalid = false;
while ~isvalid
    rows = randperm(gridSize(1),3);
    cols = randperm(gridSize(2),3);
    s0 = [rows' cols'];
    if all(~all(s0(1,:) == obsMat,2)) && all(~all(s0(2,:) == obsMat,2)) && all(~all(s0(3,:) == obsMat,2))
        isvalid = true;
    end
end

% initial grid
g0 = zeros(12,12);

% obstacle cells
for idx = 1:size(obsMat,1)
    r = obsMat(idx,1);
    c = obsMat(idx,2);
    g0(r,c) = 1.0;
end

% robot cells
g0(s0(1,1),s0(1,2)) = 0.25;
g0(s0(2,1),s0(2,2)) = 0.50;
g0(s0(3,1),s0(3,2)) = 0.75;

in = setVariable(in, 's0', s0);
in = setVariable(in, 'g0', g0);
end