Agents

example.cpp

#include "environment.hpp"
#include <iostream>
#include "NaiveAgent.hpp"

using namespace gym;

int main(int argc, char* argv[])
{
  const std::string environment = "SpaceInvaders-v0";
  const std::string host = "kurg.org";
  const std::string port = "4040";

  double totalReward = 0;
  size_t totalSteps = 0;

  Environment env(host, port, environment);
  env.compression(9);
  env.monitor.start("./dummy/", true, true);

  env.reset();
  env.render();

  // while (1)
  // {
  //   arma::mat action = env.action_space.sample();
  //   std::cout << "action: \n" << action << std::endl;

  //   env.step(action);

  //   totalReward += env.reward;
  //   totalSteps += 1;

  //   if (env.done)
  //   {
  //     break;
  //   }

  //   std::cout << "Current step: " << totalSteps << " current reward: "
  //     << totalReward << std::endl;
  // }

  Agent agent(210, 160);

  agent.Play(env, 0.2);

  std::cout << "Instance: " << env.instance << " total steps: " << totalSteps
      << " reward: " << totalReward << std::endl;

  return 0;
}

NaiveAgent.hpp

#include <cmath>
#include <mlpack/core.hpp>
#include <mlpack/core/optimizers/sgd/sgd.hpp>
#include <mlpack/methods/ann/layer/layer.hpp>
#include <mlpack/methods/ann/ffn.hpp>
#include <mlpack/methods/ann/layer/linear.hpp>
#include <mlpack/methods/ann/layer/dropout.hpp>
#include <mlpack/methods/ann/layer/leaky_relu.hpp>
#include <mlpack/methods/ann/layer/convolution.hpp>

using namespace mlpack;
using namespace optimization;
using namespace ann;
using namespace gym;

class Agent{
 public:
  FFN<> model;
  
  Agent(size_t inputW, size_t inputH){
    // 1 stream to 10 streams of depth, filter 5x5
    model.Add<Convolution<>>(1, 10, 5, 5, 1, 1, 0, 0, inputW, inputH);
    model.Add<LeakyReLU<>>();
    
    // 10 depth to 15 depth, filter 5x5, strides of 2
    model.Add<Convolution<>>(10, 15, 5, 5, 2, 2);
    model.Add<LeakyReLU<>>();

    model.Add<Linear<>>(115140, 700);
    model.Add<LeakyReLU<>>();
    
    // Fully Connected from 700 -> 3 nodes
    model.Add<Linear<>>(700, 3);
  }

  void Play(Environment& env, double explore){
    arma::mat result, actionMat, frame;
    double maxRewardAction;

    double totalReward = 0;
    size_t totalSteps = 0;

    while(1){
      //Get observation
      frame = arma::vectorise(env.observation);
      
      //Predict a reward for each action in current frame
      model.Predict(frame, result);
      
      if(arma::randu() > explore){
        //Pick the action with maximum reward
        maxRewardAction = arma::mat(result.t()).index_max();
      }
      else{
        //Pick random action number
        maxRewardAction = floor(arma::randu() * 3);
      }

      //Make 1 hot vector for chosen action
      actionMat = arma::zeros<arma::mat>(1, 3);
      actionMat[maxRewardAction] = 1;

      env.step(actionMat);

      // New Matrix for correct result
      // replace current reward with correct one
      arma::mat correctResult(result);
      correctResult[maxRewardAction] = env.reward;

      // Optimize for the correct result given same frame

      std::cout << "Corrected result:";
      for(int i : correctResult){
        std::cout << i << " ";
      }

      model.Train(std::move(frame), std::move(correctResult));
      
      if(env.done){
        break;
      }

      totalReward += env.reward;
      totalSteps += 1;

      std::cout << "Current step: " << totalSteps << " current reward: "
      << totalReward << std::endl;
    }
  }

};

/home/trion/Projects/GithubContrib/gym_tcp_api/cpp/NaiveAgent.hpp:52:34:   required from here
/usr/local/include/mlpack/methods/ann/visitor/output_parameter_visitor_impl.hpp:25:33: error: invalid use of incomplete type ‘class mlpack::ann::AddMerge<arma::Mat<double>, arma::Mat<double> >’
   return layer->OutputParameter();

There are several instances of these errors, only the layer name is changed everytime, error log is very large so just pasted the first one. Should I post complete?

How do you build the project? Do you link against mlpack?

The naiveagent.hpp is in the cpp folder of gym_tcp_api
I can agent.Play() in example.cpp

So in build folder, cmake .. & make > log.txt 2>&1

So you didn't modified the CMakeList file?

Nope

So the provided CMakeList file doesn't link against mlpack, but that's necessary to use mlpack functions. Take a look at this gist and the comments: https://gist.github.com/zoq/1810dcd8261078fe35dd118d454f229b

Some comments:

• Per default, the FFN class uses negative log likelihood as output criterion, which expects a single target for each input, where the target is the index of the class (first class has index 1). Also, it is expected that the input contains log-probabilities of each class. I guess what you like to do is to minimize the mean squared error or something similar, if you do you don't have to change the output just the output layer:

FFN<MeanSquaredError<>> model;

• The input is a 3rd order tensor right (RGB image), so the input size is 210, 160 as specified but the input size of the first Convolution layer is 3 and not 1:

model.Add<Convolution<>>(3, 10, 5, 5, 1, 1, 0, 0, inputW, inputH);

• I haven't really thought about this, but if you use stride 2 and kernel size 1 shouldn't you use pad = 1?

model.Add<Convolution<>>(10, 2, 5, 5, 2, 2, 1, 1);

So at the end it should look like:

FFN<MeanSquaredError<>> model;

Agent(size_t inputW, size_t inputH){
  // 1 stream to 10 streams of depth, filter 5x5
  model.Add<Convolution<>>(3, 10, 5, 5, 1, 1, 0, 0, inputW, inputH);
  model.Add<LeakyReLU<>>();
  
  // 10 depth to 15 depth, filter 5x5, strides of 2
  model.Add<Convolution<>>(10, 2, 5, 5, 2, 2, 1, 1);
  model.Add<LeakyReLU<>>();

  model.Add<Linear<>>(15708, 700);
  model.Add<LeakyReLU<>>();
  
  // Fully Connected from 700 -> 3 nodes
  model.Add<Linear<>>(700, 3);
}

Let me know if that was helpful.