mit.js

//<![CDATA[

// a few things don't have var in front of them - they update already existing variables the game needs
// Input is 7 lanes, security is 5 slots back and 6 ahead
lanesSide = 2;
patchesAhead = 11;
patchesBehind = 6;
trainIterations = 10000;

var num_inputs = (lanesSide * 2 + 1) * (patchesAhead + patchesBehind);
var num_actions = 5;
var temporal_window = 192;
var network_size = num_inputs * temporal_window + num_actions * temporal_window + num_inputs;

var layer_defs = [];
layer_defs.push({type: 'input', out_sx: 1, out_sy: 1, out_depth: network_size});
layer_defs.push({type: 'fc', num_neurons: 170, activation: 'relu' });
layer_defs.push({type: 'fc', num_neurons: 75, activation: 'relu' });
layer_defs.push({type: 'fc', num_neurons: 30, activation: 'relu' });
layer_defs.push({type: 'regression', num_neurons: num_actions });

var tdtrainer_options = {
    learning_rate: 0.001,
    momentum: 0.0,
    batch_size: 64,
    l2_decay: 0.01
};

var opt = {};
opt.temporal_window = temporal_window;
opt.experience_size = 6000;
opt.start_learn_threshold = 500;
opt.gamma = 0.7;
opt.learning_steps_burnin = 5000;
opt.epsilon_min = 0.05;
opt.epsilon_test_time = 0.01;
opt.layer_defs = layer_defs;
opt.random_action_distribution = [0.2, 0.25, 0.05, 0.25, 0.25];
opt.tdtrainer_options = tdtrainer_options;

brain = new deepqlearn.Brain(num_inputs, num_actions, opt);

learn = function (state, lastReward) {
    brain.backward(lastReward);
    var action = brain.forward(state);

    draw_net();
    draw_stats();

    return action;
}

//]]>