MNISTDigitDLRERPOMDP.cpp

Go to the documentation of this file.

#include <application/MNISTDigitDLRERPOMDP.hpp>

#include <limits>
#include <utils/RandomGenerator.hpp>


namespace mic {
namespace application {

void RegisterApplication (void) {
    REGISTER_APPLICATION(mic::application::MNISTDigitDLRERPOMDP);
}


MNISTDigitDLRERPOMDP::MNISTDigitDLRERPOMDP(std::string node_name_) : OpenGLEpisodicApplication(node_name_),
        saccadic_path(new std::vector <mic::types::Position2D>()),
        step_reward("step_reward", 0.0),
        discount_rate("discount_rate", 0.9),
        learning_rate("learning_rate", 0.005),
        epsilon("epsilon", 0.1),
        step_limit("step_limit",0),
        statistics_filename("statistics_filename","mnist_digit_drl_er_statistics.csv"),
        mlnn_filename("mlnn_filename", "mnist_digit_drl_er_mlnn.txt"),
        mlnn_save("mlnn_save", false),
        mlnn_load("mlnn_load", false),
        experiences(10000,1)
    {
    // Register properties - so their values can be overridden (read from the configuration file).
    registerProperty(step_reward);
    registerProperty(discount_rate);
    registerProperty(learning_rate);
    registerProperty(epsilon);
    registerProperty(step_limit);
    registerProperty(statistics_filename);
    registerProperty(mlnn_filename);
    registerProperty(mlnn_save);
    registerProperty(mlnn_load);

    LOG(LINFO) << "Properties registered";
}


MNISTDigitDLRERPOMDP::~MNISTDigitDLRERPOMDP() {
    delete(w_chart);
    delete(wmd_environment);
    delete(wmd_observation);    
}


void MNISTDigitDLRERPOMDP::initialize(int argc, char* argv[]) {
    // Initialize GLUT! :]
    VGL_MANAGER->initializeGLUT(argc, argv);

    collector_ptr = std::make_shared < mic::utils::DataCollector<std::string, float> >( );
    // Add containers to collector.
    collector_ptr->createContainer("path_length_episode",  mic::types::color_rgba(0, 255, 0, 180));
    collector_ptr->createContainer("path_length_average", mic::types::color_rgba(255, 255, 0, 180));
    collector_ptr->createContainer("path_length_optimal", mic::types::color_rgba(255, 255, 255, 180));
    collector_ptr->createContainer("path_length_diff", mic::types::color_rgba(255, 0, 0, 180));

    sum_of_iterations = 0;

    // Create the visualization windows - must be created in the same, main thread :]
    w_chart = new WindowCollectorChart<float>("MNISTDigitDLRERPOMDP", 256, 512, 0, 0);
    w_chart->setDataCollectorPtr(collector_ptr);

}

void MNISTDigitDLRERPOMDP::initializePropertyDependentVariables() {
    // Create windows for the visualization of the whole environment and a single observation.
    wmd_environment = new WindowMNISTDigit("Environment", env.getEnvironmentHeight()*20,env.getEnvironmentWidth()*20, 0, 316);
    wmd_observation = new WindowMNISTDigit("Observation", env.getObservationHeight()*20,env.getObservationWidth()*20, env.getEnvironmentWidth()*20, 316);


    // Hardcode batchsize - for fastening the display!
    batch_size = env.getObservationWidth() * env.getObservationHeight();

    // Try to load neural network from file.
    if ((mlnn_load) && (neural_net.load(mlnn_filename))) {
        // Do nothing ;)
    } else {
        // Create a simple neural network.
        // gridworld wxhx4 -> 100 -> 4 -> regression!.
        neural_net.pushLayer(new Linear<float>((size_t) env.getObservationSize(), 250));
        neural_net.pushLayer(new ReLU<float>(250));
        neural_net.pushLayer(new Linear<float>(250, 100));
        neural_net.pushLayer(new ReLU<float>(100));
        neural_net.pushLayer(new Linear<float>(100, 4));

        // Set batch size.
        neural_net.resizeBatch(batch_size);
        // Change optimization function from default GradientDescent to Adam.
        neural_net.setOptimization<mic::neural_nets::optimization::Adam<float> >();
        // Set loss function -> regression!
        neural_net.setLoss <mic::neural_nets::loss::SquaredErrorLoss<float> >();

        LOG(LINFO) << "Generated new neural network";
    }//: else

    // Set batch size in experience replay memory.
    experiences.setBatchSize(batch_size);

    // Set displayed matrix pointers.
    wmd_environment->setDigitPointer(env.getEnvironment());
    wmd_environment->setPathPointer(saccadic_path);
    wmd_observation->setDigitPointer(env.getObservation());

}


void MNISTDigitDLRERPOMDP::startNewEpisode() {
    LOG(LSTATUS) << "Starting new episode " << episode;

    // Generate the gridworld (and move player to initial position).
    env.initializeEnvironment();
    saccadic_path->clear();
    // Add first, initial position to  to saccadic path.
    saccadic_path->push_back(env.getAgentPosition());

    /*LOG(LNOTICE) << "Network responses: \n" <<  streamNetworkResponseTable();
    LOG(LNOTICE) << "Observation: \n"  << env.observationToString();
    LOG(LNOTICE) << "Environment: \n" << env.environmentToString();*/
    // Do not forget to get the current observation!
    env.getObservation();
}


void MNISTDigitDLRERPOMDP::finishCurrentEpisode() {
    LOG(LTRACE) << "End current episode";

    sum_of_iterations += iteration -1; // -1 is the fix related to moving the terminal condition to the front of step!

    // Add variables to container.
    collector_ptr->addDataToContainer("path_length_episode",(iteration -1));
    collector_ptr->addDataToContainer("path_length_average",(float)sum_of_iterations/episode);
    collector_ptr->addDataToContainer("path_length_optimal", (float)env.optimalPathLength());
    collector_ptr->addDataToContainer("path_length_diff", (float)(iteration -1 - env.optimalPathLength()));


    // Export reward "convergence" diagram.
    collector_ptr->exportDataToCsv(statistics_filename);

    // Save nn to file.
    if (mlnn_save && (episode %10))
        neural_net.save(mlnn_filename);
}


std::string MNISTDigitDLRERPOMDP::streamNetworkResponseTable() {
    LOG(LTRACE) << "streamNetworkResponseTable()";
    std::string rewards_table;
    std::string actions_table;

    // Remember the current state i.e. player position.
    mic::types::Position2D current_player_pos_t = env.getAgentPosition();

    // Create new matrices for batches of inputs and targets.
    MatrixXfPtr inputs_batch(new MatrixXf(env.getObservationSize(), batch_size));

    // Assume that the batch_size = grid_env.getWidth() * grid_env.getHeight()
    assert(env.getObservationWidth()*env.getObservationHeight() == batch_size);


    size_t dx = (env.getObservationWidth()-1)/2;
    size_t dy = (env.getObservationHeight()-1)/2;
    mic::types::Position2D p = env.getAgentPosition();

    // Copy data.
    for (long oy=0, ey=(p.y-dy); oy<(long)env.getObservationHeight(); oy++, ey++){
        for (long ox=0, ex=(p.x-dx); ox<(long)env.getObservationWidth(); ox++, ex++) {

            // Move the player to given state - disregarding whether it was successful or not, answers for walls/positions outside of the gridworld do not interes us anyway...
            if (!env.moveAgentToPosition(Position2D(ex,ey)))
                LOG(LDEBUG) << "Failed!"; //... but still we can live with that... ;)
            // Encode the current state.
            mic::types::MatrixXfPtr encoded_state = env.encodeObservation();
            // Add to batch.
            inputs_batch->col(oy*env.getObservationWidth()+ox) = encoded_state->col(0);
        }//: for x
    }//: for y

    // Get rewards for the whole batch.
    neural_net.forward(inputs_batch);
    // Get predictions for all those states - there is no need to create a copy.
    MatrixXfPtr predicted_batch = neural_net.getPredictions();


    rewards_table += "Action values:\n";
    actions_table += "Best actions:\n";
    // Generate all possible states and all possible rewards.
    for (long oy=0, ey=(p.y-dy); oy<(long)env.getObservationHeight(); oy++, ey++){
        rewards_table += "| ";
        actions_table += "| ";
        for (long ox=0, ex=(p.x-dx); ox<(long)env.getObservationWidth(); ox++, ex++) {
            float bestqval = -std::numeric_limits<float>::infinity();
            size_t best_action = -1;
            for (size_t a=0; a<4; a++) {
                float qval = (*predicted_batch)(a, oy*env.getObservationWidth()+ox);

                rewards_table += std::to_string(qval);
                if (a==3)
                    rewards_table += " | ";
                else
                    rewards_table += " , ";

                // Remember the best value.
                if (env.isStateAllowed(ex,ey) && (!env.isStateTerminal(ex,ey)) && env.isActionAllowed(ex,ey,a) && (qval > bestqval)){
                    bestqval = qval;
                    best_action = a;
                }//: if

            }//: for a(ctions)
            switch(best_action){
                case 0 : actions_table += "N | "; break;
                case 1 : actions_table += "E | "; break;
                case 2 : actions_table += "S | "; break;
                case 3 : actions_table += "W | "; break;
                default: actions_table += "- | ";
            }//: switch

        }//: for x
        rewards_table += "\n";
        actions_table += "\n";
    }//: for y

    // Move player to previous position.
    env.moveAgentToPosition(current_player_pos_t);

    return rewards_table + actions_table;
}



float MNISTDigitDLRERPOMDP::computeBestValueForGivenStateAndPredictions(mic::types::Position2D player_position_, float* predictions_){
    LOG(LTRACE) << "computeBestValueForGivenState()";
    float best_qvalue = -std::numeric_limits<float>::infinity();

    // Create a list of possible actions.
    std::vector<mic::types::NESWAction> actions;
    actions.push_back(A_NORTH);
    actions.push_back(A_EAST);
    actions.push_back(A_SOUTH);
    actions.push_back(A_WEST);

    for(mic::types::NESWAction action : actions) {
        // .. and find the value of the best allowed action.
        if(env.isActionAllowed(player_position_, action)) {
            float qvalue = predictions_[(size_t)action.getType()];
            if (qvalue > best_qvalue)
                best_qvalue = qvalue;
        }//if is allowed
    }//: for

    return best_qvalue;
}


mic::types::MatrixXfPtr MNISTDigitDLRERPOMDP::getPredictedRewardsForGivenState(mic::types::Position2D player_position_) {
    LOG(LTRACE) << "getPredictedRewardsForGivenState()";
    // Remember the current state i.e. player position.
    mic::types::Position2D current_player_pos_t = env.getAgentPosition();

    // Move the player to given state.
    env.moveAgentToPosition(player_position_);

    // Encode the current state.
    mic::types::MatrixXfPtr encoded_state = env.encodeObservation();

    // Create NEW matrix for the inputs batch.
    MatrixXfPtr inputs_batch(new MatrixXf(env.getObservationSize(), batch_size));
    inputs_batch->setZero();

    // Set the first input - only this one interests us.
    inputs_batch->col(0) = encoded_state->col(0);

    //LOG(LERROR) << "Getting predictions for input batch:\n" <<inputs_batch->transpose();

    // Pass the data and get predictions.
    neural_net.forward(inputs_batch);

    MatrixXfPtr predictions_batch = neural_net.getPredictions();

    //LOG(LERROR) << "Resulting predictions batch:\n" << predictions_batch->transpose();

    // Get the first prediction only.
    MatrixXfPtr predictions_sample(new MatrixXf(4, 1));
    predictions_sample->col(0) = predictions_batch->col(0);

    //LOG(LERROR) << "Returned predictions sample:\n" << predictions_sample->transpose();

    // Move player to previous position.
    env.moveAgentToPosition(current_player_pos_t);

    // Return the predictions.
    return predictions_sample;
}

mic::types::NESWAction MNISTDigitDLRERPOMDP::selectBestActionForGivenState(mic::types::Position2D player_position_){
    LOG(LTRACE) << "selectBestAction";

    // Greedy methods - returns the index of element with greatest value.
    mic::types::NESWAction best_action = A_RANDOM;
    float best_qvalue = -std::numeric_limits<float>::infinity();

    // Create a list of possible actions.
    std::vector<mic::types::NESWAction> actions;
    actions.push_back(A_NORTH);
    actions.push_back(A_EAST);
    actions.push_back(A_SOUTH);
    actions.push_back(A_WEST);

    // Check the results of actions one by one... (there is no need to create a separate copy of predictions)
    MatrixXfPtr predictions_sample = getPredictedRewardsForGivenState(player_position_);
    //LOG(LERROR) << "Selecting action from predictions:\n" << predictions_sample->transpose();
    float* pred = predictions_sample->data();

    for(size_t a=0; a<4; a++) {
        // Find the best action allowed.
        if(env.isActionAllowed(player_position_, mic::types::NESWAction((mic::types::NESW)a))) {
            float qvalue = pred[a];
            if (qvalue > best_qvalue){
                best_qvalue = qvalue;
                best_action.setAction((mic::types::NESW)a);
            }
        }//if is allowed
    }//: for

    return best_action;
}

bool MNISTDigitDLRERPOMDP::performSingleStep() {
    LOG(LSTATUS) << "Episode "<< episode << ": step " << iteration << "";

    // Get player pos at time t.
    mic::types::Position2D player_pos_t= env.getAgentPosition();
    LOG(LINFO) << "Agent position at state t: " << player_pos_t;

    // Check whether state t is terminal - finish the episode.
    if(env.isStateTerminal(player_pos_t))
        return false;

    // TMP!
    double  nn_weight_decay = 0;

    // Select the action.
    mic::types::NESWAction action;
    //action = A_NORTH;
    double eps = (double)epsilon;
    if ((double)epsilon < 0)
        eps = 1.0/(1.0+sqrt(episode));
    if (eps < 0.1)
        eps = 0.1;
    LOG(LDEBUG) << "eps = " << eps;
    bool random = false;

    // Epsilon-greedy action selection.
    if (RAN_GEN->uniRandReal() > eps){
        // Select best action.
        action = selectBestActionForGivenState(player_pos_t);
    } else {
        // Random action.
        action = A_RANDOM;
        random = true;
    }//: if

    // Execute action - do not monitor the success.
    env.moveAgent(action);

    // Get new state s(t+1).
    mic::types::Position2D player_pos_t_prim = env.getAgentPosition();
    LOG(LINFO) << "Agent position at t+1: " << player_pos_t_prim << " after performing the action = " << action << ((random) ? " [Random]" : "");

    // Add this position to  to saccadic path.
    saccadic_path->push_back(player_pos_t_prim);

    // Collect the experience.
    SpatialExperiencePtr exp(new SpatialExperience(player_pos_t, action, player_pos_t_prim));
    // Create an empty matrix for rewards - this will be recalculated each time the experience will be replayed anyway.
    MatrixXfPtr rewards (new MatrixXf(4 , batch_size));
    // Add experience to experience table.
    experiences.add(exp, rewards);


    // Deep Q learning - train network with random sample from the experience memory.
    if (experiences.size() >= 2*batch_size) {
        // Create new matrices for batches of inputs and targets.
        MatrixXfPtr inputs_t_batch(new MatrixXf(env.getObservationSize(), batch_size));
        MatrixXfPtr inputs_t_prim_batch(new MatrixXf(env.getObservationSize(), batch_size));
        MatrixXfPtr targets_t_batch(new MatrixXf(4, batch_size));

        // Get the random batch.
        SpatialExperienceBatch geb = experiences.getRandomBatch();

        // Debug purposes.
        geb.setNextSampleIndex(0);
        for (size_t i=0; i<batch_size; i++) {
            SpatialExperienceSample ges = geb.getNextSample();
            SpatialExperiencePtr ge_ptr = ges.data();
            LOG(LDEBUG) << "Training sample : " << ge_ptr->s_t << " -> " << ge_ptr->a_t << " -> " << ge_ptr->s_t_prim;
        }//: for

        // Iterate through samples and create inputs_t_batch.
        for (size_t i=0; i<batch_size; i++) {
            SpatialExperienceSample ges = geb.getNextSample();
            SpatialExperiencePtr ge_ptr = ges.data();

            // Replay the experience.
            // "Simulate" moving player to position from state/time (t).
            env.moveAgentToPosition(ge_ptr->s_t);
            // Encode the state at time (t).
            mic::types::MatrixXfPtr encoded_state_t = env.encodeObservation();
            //float* state = encoded_state_t->data();

            // Copy the encoded state to inputs batch.
            inputs_t_batch->col(i) = encoded_state_t->col(0);
        }// for samples.

        // Get network responses.
        neural_net.forward(inputs_t_batch);
        // Get predictions for all those states...
        MatrixXfPtr predictions_t_batch = neural_net.getPredictions();
        // ... and copy them to reward pointer - a container which we will modify.
        (*targets_t_batch) = (*predictions_t_batch);

        // Iterate through samples and create inputs_t_prim_batch.
        geb.setNextSampleIndex(0);
        for (size_t i=0; i<batch_size; i++) {
            SpatialExperienceSample ges = geb.getNextSample();
            SpatialExperiencePtr ge_ptr = ges.data();

            // Replay the experience.
            // "Simulate" moving player to position from state/time (t+1).
            env.moveAgentToPosition(ge_ptr->s_t_prim);
            // Encode the state at time (t+1).
            mic::types::MatrixXfPtr encoded_state_t = env.encodeObservation();
            //float* state = encoded_state_t->data();

            // Copy the encoded state to inputs batch.
            inputs_t_prim_batch->col(i) = encoded_state_t->col(0);
        }// for samples.

        // Get network responses.
        neural_net.forward(inputs_t_prim_batch);
        // Get predictions for all those states...
        MatrixXfPtr predictions_t_prim_batch = neural_net.getPredictions();

        // Calculate the rewards, one by one.
        // Iterate through samples and create inputs_t_prim_batch.
        geb.setNextSampleIndex(0);
        for (size_t i=0; i<batch_size; i++) {
            SpatialExperienceSample ges = geb.getNextSample();
            SpatialExperiencePtr ge_ptr = ges.data();

            if (ge_ptr->s_t == ge_ptr->s_t_prim) {
                // The move was not possible! Learn that as well.
                (*targets_t_batch)((size_t)ge_ptr->a_t.getType(), i) = 3*step_reward;
            } else if(env.isStateTerminal(ge_ptr->s_t_prim)) {
                // The position at (t+1) state appears to be terminal - learn the reward.
                (*targets_t_batch)((size_t)ge_ptr->a_t.getType(), i) = env.getStateReward(ge_ptr->s_t_prim);
            } else {
                MatrixXfPtr preds_t_prim (new MatrixXf(4, 1));
                preds_t_prim->col(0) = predictions_t_prim_batch->col(i);
                // Get best value for the NEXT state - position from (t+1) state.
                float max_q_st_prim_at_prim = computeBestValueForGivenStateAndPredictions(ge_ptr->s_t_prim, preds_t_prim->data());
                // If next state best value is finite.
                // Update running average for given action - Deep Q learning!
                if (std::isfinite(max_q_st_prim_at_prim))
                    (*targets_t_batch)((size_t)ge_ptr->a_t.getType(), i) = step_reward + discount_rate*max_q_st_prim_at_prim;
            }//: else

        }//: for

        LOG(LDEBUG) <<"Inputs batch:\n" << inputs_t_batch->transpose();
        LOG(LDEBUG) <<"Targets batch:\n" << targets_t_batch->transpose();

        // Perform the Deep-Q-learning.
        //LOG(LDEBUG) << "Network responses before training:" << std::endl << streamNetworkResponseTable();

        // Train network with rewards.
        float loss = neural_net.train (inputs_t_batch, targets_t_batch, learning_rate, nn_weight_decay);
        LOG(LDEBUG) << "Training loss:" << loss;

        //LOG(LDEBUG) << "Network responses after training:" << std::endl << streamNetworkResponseTable();

        // Finish the replay: move the player to REAL, CURRENT POSITION.
        env.moveAgentToPosition(player_pos_t_prim);
    }//: if enough experiences
    else
        LOG(LWARNING) << "Not enough samples in the experience replay memory!";

    LOG(LNOTICE) << "Network responses: \n" << streamNetworkResponseTable();
    LOG(LNOTICE) << "Observation: \n"  << env.observationToString();
    LOG(LNOTICE) << "Environment: \n"  << env.environmentToString();
    // Do not forget to get the current observation!
    env.getObservation();

    // Check whether we reached maximum number of iterations.
    if ((step_limit>0) && (iteration > (size_t)step_limit))
        return false;

    return true;
}

} /* namespace application */
} /* namespace mic */