blob: e96a5179926dcd93f62b9b000e21d9cf5485e72e [file] [log] [blame]
/*
* Copyright (C) 2015-2017 Apple Inc. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY APPLE INC. AND ITS CONTRIBUTORS ``AS IS''
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
* THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR ITS CONTRIBUTORS
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
* THE POSSIBILITY OF SUCH DAMAGE.
*/
SimpleKalmanEstimator = Utilities.createSubclass(Experiment,
function(processError, measurementError) {
Experiment.call(this, false);
var error = .5 * (Math.sqrt(processError * processError + 4 * processError * measurementError) - processError);
this._gain = error / (error + measurementError);
}, {
sample: function(newMeasurement)
{
if (!this._initialized) {
this._initialized = true;
this.estimate = newMeasurement;
return;
}
this.estimate = this.estimate + this._gain * (newMeasurement - this.estimate);
},
reset: function()
{
Experiment.prototype.reset.call(this);
this._initialized = false;
this.estimate = 0;
}
});
PIDController = Utilities.createClass(
function(ysp)
{
this._ysp = ysp;
this._out = 0;
this._Kp = 0;
this._stage = PIDController.stages.WARMING;
this._eold = 0;
this._I = 0;
}, {
// Determines whether the current y is
// before ysp => (below ysp if ysp > y0) || (above ysp if ysp < y0)
// after ysp => (above ysp if ysp > y0) || (below ysp if ysp < y0)
_yPosition: function(y)
{
return (y < this._ysp) == (this._y0 < this._ysp)
? PIDController.yPositions.BEFORE_SETPOINT
: PIDController.yPositions.AFTER_SETPOINT;
},
// Calculate the ultimate distance from y0 after time t. We want to move very
// slowly at the beginning to see how adding few items to the test can affect
// its output. The complexity of a single item might be big enough to keep the
// proportional gain very small but achieves the desired progress. But if y does
// not change significantly after adding few items, that means we need a much
// bigger gain. So we need to move over a cubic curve which increases very
// slowly with small t values but moves very fast with larger t values.
// The basic formula is: y = t^3
// Change the formula to reach y=1 after 1000 ms: y = (t/1000)^3
// Change the formula to reach y=(ysp - y0) after 1000 ms: y = (ysp - y0) * (t/1000)^3
_distanceUltimate: function(t)
{
return (this._ysp - this._y0) * Math.pow(t / 1000, 3);
},
// Calculates the distance of y relative to y0. It also ensures we do not return
// zero by returning a epsilon value in the same direction as ultimate distance.
_distance: function(y, du)
{
const epsilon = 0.0001;
var d = y - this._y0;
return du < 0 ? Math.min(d, -epsilon) : Math.max(d, epsilon);
},
// Decides how much the proportional gain should be increased during the manual
// gain stage. We choose to use the ratio of the ultimate distance to the current
// distance as an indication of how much the system is responsive. We want
// to keep the increment under control so it does not cause the system instability
// So we choose to take the natural logarithm of this ratio.
_gainIncrement: function(t, y, e)
{
var du = this._distanceUltimate(t);
var d = this._distance(y, du);
return Math.log(du / d) * 0.1;
},
// Update the stage of the controller based on its current stage and the system output
_updateStage: function(y)
{
var yPosition = this._yPosition(y);
switch (this._stage) {
case PIDController.stages.WARMING:
if (yPosition == PIDController.yPositions.AFTER_SETPOINT)
this._stage = PIDController.stages.OVERSHOOT;
break;
case PIDController.stages.OVERSHOOT:
if (yPosition == PIDController.yPositions.BEFORE_SETPOINT)
this._stage = PIDController.stages.UNDERSHOOT;
break;
case PIDController.stages.UNDERSHOOT:
if (yPosition == PIDController.yPositions.AFTER_SETPOINT)
this._stage = PIDController.stages.SATURATE;
break;
}
},
// Manual tuning is used before calculating the PID controller gains.
_tuneP: function(e)
{
// The output is the proportional term only.
return this._Kp * e;
},
// PID tuning function. Kp, Ti and Td were already calculated
_tunePID: function(h, y, e)
{
// Proportional term.
var P = this._Kp * e;
// Integral term is the area under the curve starting from the beginning
// till the current time.
this._I += (this._Kp / this._Ti) * ((e + this._eold) / 2) * h;
// Derivative term is the slope of the curve at the current time.
var D = (this._Kp * this._Td) * (e - this._eold) / h;
// The ouput is a PID function.
return P + this._I + D;
},
// Apply different strategies for the tuning based on the stage of the controller.
_tune: function(t, h, y, e)
{
switch (this._stage) {
case PIDController.stages.WARMING:
// This is the first stage of the Zieglerâ€Nichols method. It increments
// the proportional gain till the system output passes the set-point value.
if (typeof this._y0 == "undefined") {
// This is the first time a tuning value is required. We want the test
// to add only one item. So we need to return -1 which forces us to
// choose the initial value of Kp to be = -1 / e
this._y0 = y;
this._Kp = -1 / e;
} else {
// Keep incrementing the Kp as long as we have not reached the
// set-point yet
this._Kp += this._gainIncrement(t, y, e);
}
return this._tuneP(e);
case PIDController.stages.OVERSHOOT:
// This is the second stage of the Zieglerâ€Nichols method. It measures the
// oscillation period.
if (typeof this._t0 == "undefined") {
// t is the time of the begining of the first overshot
this._t0 = t;
this._Kp /= 2;
}
return this._tuneP(e);
case PIDController.stages.UNDERSHOOT:
// This is the end of the Zieglerâ€Nichols method. We need to calculate the
// integral and derivative periods.
if (typeof this._Ti == "undefined") {
// t is the time of the end of the first overshot
var Tu = t - this._t0;
// Calculate the system parameters from Kp and Tu assuming
// a "some overshoot" control type. See:
// https://en.wikipedia.org/wiki/Ziegler%E2%80%93Nichols_method
this._Ti = Tu / 2;
this._Td = Tu / 3;
this._Kp = 0.33 * this._Kp;
// Calculate the tracking time.
this._Tt = Math.sqrt(this._Ti * this._Td);
}
return this._tunePID(h, y, e);
case PIDController.stages.SATURATE:
return this._tunePID(h, y, e);
}
return 0;
},
// Ensures the system does not fluctuates.
_saturate: function(v, e)
{
var u = v;
switch (this._stage) {
case PIDController.stages.OVERSHOOT:
case PIDController.stages.UNDERSHOOT:
// Calculate the min-max values of the saturation actuator.
if (typeof this._min == "undefined")
this._min = this._max = this._out;
else {
this._min = Math.min(this._min, this._out);
this._max = Math.max(this._max, this._out);
}
break;
case PIDController.stages.SATURATE:
const limitPercentage = 0.90;
var min = this._min > 0 ? Math.min(this._min, this._max * limitPercentage) : this._min;
var max = this._max < 0 ? Math.max(this._max, this._min * limitPercentage) : this._max;
var out = this._out + u;
// Clip the controller output to the min-max values
out = Math.max(Math.min(max, out), min);
u = out - this._out;
// Apply the back-calculation and tracking
if (u != v)
u += (this._Kp * this._Tt / this._Ti) * e;
break;
}
this._out += u;
return u;
},
// Called from the benchmark to tune its test. It uses Ziegler-Nichols method
// to calculate the controller parameters. It then returns a PID tuning value.
tune: function(t, h, y)
{
this._updateStage(y);
// Current error.
var e = this._ysp - y;
var v = this._tune(t, h, y, e);
// Save e for the next call.
this._eold = e;
// Apply back-calculation and tracking to avoid integrator windup
return this._saturate(v, e);
}
});
Utilities.extendObject(PIDController, {
// This enum will be used to tell whether the system output (or the controller input)
// is moving towards the set-point or away from it.
yPositions: {
BEFORE_SETPOINT: 0,
AFTER_SETPOINT: 1
},
// The Ziegler-Nichols method for is used tuning the PID controller. The workflow of
// the tuning is split into four stages. The first two stages determine the values
// of the PID controller gains. During these two stages we return the proportional
// term only. The third stage is used to determine the min-max values of the
// saturation actuator. In the last stage back-calculation and tracking are applied
// to avoid integrator windup. During the last two stages, we return a PID control
// value.
stages: {
WARMING: 0, // Increase the value of the Kp until the system output reaches ysp.
OVERSHOOT: 1, // Measure the oscillation period and the overshoot value
UNDERSHOOT: 2, // Return PID value and measure the undershoot value
SATURATE: 3 // Return PID value and apply back-calculation and tracking.
}
});