This is a working example of how to use the DQNSolver-Class in your Project. For further information on:
- how the solver is constructed,
- what all the single values stand for and
- how the solver can action on it's environment,
please also have a look here.
To see a live example visit learning-agents.
// TypeScript
import { DQNSolver, DQNOpt, DQNEnv } from 'reinforce-js';
// JavaScript
// const DQNSolver = require('reinforce-js').DQNSolver;
// const DQNEnv = require('reinforce-js').DQNEnv;
// const DQNOpt = require('reinforce-js').DQNOpt;
const width = 400;
const height = 400;
const numberOfStates = 20;
const numberOfActions = 4;
const env = new DQNEnv(width, height, numberOfStates, numberOfActions);
const opt = new DQNOpt();
opt.setTrainingMode(true);
opt.setNumberOfHiddenUnits([100]); // mind the array here, currently only one layer supported! Preparation for DNN in progress...
opt.setEpsilonDecay(1.0, 0.1, 1e6);
opt.setEpsilon(0.05);
opt.setGamma(0.9);
opt.setAlpha(0.005);
opt.setLossClipping(true);
opt.setLossClamp(1.0);
opt.setRewardClipping(true);
opt.setRewardClamp(1.0);
opt.setExperienceSize(1e6);
opt.setReplayInterval(5);
opt.setReplaySteps(5);
/*
Outfit solver with environment complexity and specs.
After configuration it's ready to train its untrained Q-Network and learn from SARSA experiences.
*/
const dqnSolver = new DQNSolver(env, opt);
/*
Determine a state, e.g.:
*/
const state = [ /* Array with numerical values and length of 20 as configured via numberOfStates */ ];
/*
Now inject state and receive the preferred action as index from 0 to 3 as configured via numberOfActions.
*/
const action = dqnSolver.decide(state);
/*
Now calculate some Reward and let the Solver learn from it, e.g.:
*/
const reward = 0.9;
dqnSolver.learn(reward);