reinforcement-learning

Easy reinforcement learning using tensorflow.js

• [X] Deep Q Network (DQN)

• [ ] Genetic Algorithim (GA)

• [ ] Examples

• [ ] Training Dashboard (Tensorboard)

Require

const  rl = require('reinforcement-learning');

DQN

Parameters

• arch - Architechture of the neural network

• epsilon - % of actions that should be taken randomly for exploration

• epsilonDecay - Epsilon will be multiplied by this amount every episode

• replayMemorySize - Amount of previous steps left in memory to train on

• miniBatchSize - Batch size to fit on

• actionSpaceSize - Amount of possible actions the agent can take

• minReplaySize - Minimum amount of memories allowed for fitting

• updateTargetEvery How many episodes to wait to update the predictions network

• accuracyLookbackSize How many previous steps should be used to calculate accuracy

const  rl = require('reinforcement-learning');
let  step = 0;


let  arch = [
{inputShape:  1, units:  14, activation:  'relu'},
{units:  2, activation:  'softmax'}
];
  
function  getState(){return [0];}

function  calcReward(state, action){
// Every 100 steps end the episode
step++;
let  episodeDone = false;
if(step === 100){episodeDone = true; step = 0;}

// Two armed bandit. Agent has to learn to always pick 1
if(action === 1)return {reward:  1, newState:[0], done:  true, episodeDone};
else{ return {reward:  0, newState:[0], done:  true, episodeDone}}
}


(async()=>{
let  agent = rl.DQN({
arch,
epsilon:  0,
epsilonDecay:  0,
replayMemorySize:  1000,
miniBatchSize:  64,
actionSpaceSize:  2,
minReplaySize:  100,
updateTargetEvery:  1,
accuracyLookbackSize:  500
});

  
await  agent.learn({accuracy:  95, getState, calcReward});

  
})();

License

MIT