Saturday, August 8, 2015

Hackathon: Writing a Microsoft Band App In One Weekend - Update 3

Update 3: Wet and Wild

Data! We have Data! We have tons and tons and tons of Data!
Samples for 40 Strokes at Sprint Pace
Those same 40 strokes charted and tweaked for algorithm building

Sample Gathering

This morning I made a first pass of the Stroke Counter algorithm. I then went down to the lake and collected 10 sample groups to run against the Stroke Counter.

The sample groups included:

  • 20 Strokes at an Easy Pace
  • 5 Strokes at an Easy Pace
  • 40 Strokes at a Sprint Pace
  • 60 Strokes at an Easy Pace

Turn Simulation
  • 20 Strokes on Left Side Only
  • 10 Strokes on Left Side Only
  • 20 Strokes on Right Side Only
  • 10 Strokes on Right Side Only

Odd But Why Not
  • 20 Strokes Backwards
  • 20 Strokes Backwards

Algorithm Test Run 1

The first test run of the sample data against the initial Stroke Counter sucked...I mean it was very inaccurate. There goes trying to "guess" my way through this.

Ok, so more work was needed on the algorithm and that was my day. About 9+ hours of tweaking later and I am passing all Easy Pace tests. Tomorrow I will run a new sample session with this algorithm and see how it stands up during use in the kayak.

Improvements For Sprinting

For Sprinting or fast and strong strokes I am going to tackle the problem by looking at peak samples in trailing sample ranges. If a peak reading goes above a known Sprint High Water Mark (HWM) that is outside the normal HWM then I will apply a modifier to my HWM and LWM (Low Water Mark). Then as long as the trailing peak samples above the "Sprint HWM" I will apply a modifier to my stroke classification.

I can see that one of the trickiest problems will be handling the transition from an Easy pace to a Sprint pace. This will be interesting to solve.

Tweaking The Algorithm

So, what did I do to tweak the algorithm today?

The exporter I built into the app exports the sample data into a file with Comma Separated Values, with one reading per line. I am using the Share Sheet from my app so I can easily save my samples to Dropbox after each sample collection.

Once I got back to the house. I imported the sample groups into Numbers and created charts for each sample set.

For each chart I left the chart zoomed out on top and then I created a duplicate chart below and zoomed it in. I applied the following adjustments to my chart:
  • Series
    • Turned on Trendlines with Moving Average and tweaked the Period to find a sweet spot for the algorithm high and low water marks.
    • Turned on Error Bars with Standard Deviation to help isolate the Trendline while reviewing between the 10 sample sets.
    • Changed the Trendlines to Red and increased them to 4 pt. This helped them stand out in the charts.
  • Axis
    • I added Minor Ticks and tweaked each chart so I could better isolate the peaks and valleys of the sample.

Beginning of the Easy Pace 60 Strokes Group

Cool Charts! Now What?

I got several suggestions from friends to do machine learning, use Azure, or many other tools. For many of these, they are overkill for the initial app's stroke counter. As well, I am trying to create the initial app over this weekend so I don't have enough cycles to get the app working and learn machine learning. I did dive into machine learning a bit today and picked up some pointers to help me design the stroke counter.

I plan to take time to do further refining of the stroke counter later and look at how machine learning could help with that. For now, let's get the basics working.

So, I did the tweaking the hard way. I wrote test cases that consume my sample files and run them against the stroke counter algorithm. This has been very helpful and allowed me to tweak the algorithm to the current state (mentioned earlier) where I am correctly detecting strokes for an Easy Pace. I am also 75% correct for Sprint Pace but I want to get this closer to 100% accuracy.

Time for Pizza

It's not a Hackathon without Pizza. Calzones are on the way and then I get back to coding.

No comments: