The last post and its conclusions (in a nutshell: start the bout in slow prep and bias towards the attack; move towards fast prep and defence in the second half) seem to have generated some interest and more than a few questions. Main groups of questions were:
- What datasets did you use, and plan to use?
- How are you classifying preparations and actions, and do you have a rigorous set of definitions for each?
- Really, defense?
I’m currently laid up for at least a week tending to another quadriceps tendonitis flare-up (9 months and counting), so here, have some answers and more stats.
What datasets did you use and plan to use?
We used competitions with World Cup-grade sabre fencers, even if the event itself was not a world cup. To date the dataset is small but should grow every time new competition videos become available. Bouts used to date were:
- Asian Championships 2010 Men’s Sabre Team, Gold medal match, Korea v.s. China
- Warsaw Grand Prix 2011 Men’s Sabre Individual, Gold medal match, Won Woo Young v.s. Alexey Yakimenko
- Chicago World Cup 2013, Men’s Sabre Individual, Gold medal match, Szabo v.s. Szilagyi
- London Olympics 2012, Men’s Sabre Team, Gold medal match, Romania v.s. Korea
- Chicago World Cup 2013, Men’s Sabre Team, Gold Medal match, Russia v.s. Italy
Being as we are in Australia, we rely pretty heavily on videos from the FIE and Andrew Fischl (a.k.a CyrusofChaos) for the analysis. We do plan on wandering off to a couple of World Cups this year to grab some of our own videos as well.
How are you classifying preparations and actions, and do you have a rigorous set of definitions for each?
We’re classifying preparations and actions according the main teaching syllabus used at the Sydney Sabre Centre (aka the Codex). The terms are broadly based on those used in fencing clubs based in the USA, the UK and Australia but have been extensively revised and clarified (e.g. we make a distinction between a ‘step’ with just the front foot, and an ‘advance’ made with a front step and bringing up the back foot). When the latest version gets updated I’ll probably post the glossary to the blog, but it’s a fairly involved affair with several hundred entries.
The other part is whether the classification of actions themselves are rigorous. That’s a tougher question to answer. Most the analysis has been done by me at this stage, so at least the data are more or less consistent. The other folks who are doing the classifications are all sabre fencers who have been trained at Sydney Sabre for at least a year, so the classifications should also be broadly consistent with how I’m doing it. Upshot is that I think the data are okay, but the noise level is high and we should be careful about drawing conclusions from small statistical differences, especially before we’ve tested it on the piste.
Really, defense with fast prep?
Yes. Though our local itinerant Korean swordsman, Dong Hwan Kim, was stressing the point that we should be starting with slow prep with a bias on the attack. Ably demonstrated at a recent NSW state competition before he started messing around:
One of the notable things in this bout was how more than half the successful preparations were slow, followed by medium prep. Fast preparations made up less than 10% of the successful preparations. What’s more interesting is what the different preparations worked against: slow preparations succeeded against all other types of preparation, but fast never beat slow, and medium only worked against medium prep.
From slow preparation the most common winning action was AiT, and this combo made up about a third of all successful actions in the entire bout.
The second most common action was advance then fall short, usually from either a slow or medium prep. These occurred about a quarter of the time. Yet these combos had fairly bad odds of success – advance fall short in slow and medium prep made up about half of all unsuccessful actions as well.
Less commonly attempted actions which had an excellent success rate were advance draw cut, AoP fall short and advance back step counter-attack. Why? One possibility is that each of these actions encourage the opponent to do what you want them to do, e.g. advance back step encourages the opponent to hold their attack for the counterattack, advance (extension) induces the opponent to launch an extended attack for your draw cut, and AoP/check helps your opponent finish short. By contrast, advance fall short induces your opponent to hold their attack, which isn’t great if you’re attempting to make them fall short.
So – slow preparations and attacks work. At least if you’re a former Korean national junior team fencer up against a totally outclassed opponent (which for now is pretty much all of us).
In the next few posts I’ll start laying out some early ideas about how to do decent stats analysis on actions outside of the 4m – a harder problem because of more external factors, more scope for false positive/negatives, and greater variety of parameters – and a couple of articles on the biomechanics of sabre footwork and how we’re going to get my knee back in shape before the Athens World Cup.
I’m liking the approach here and as the data builds over time it will develop our understanding. However, I think it is important to take different perspectives. Slow preparations are the most effective in dealing with both a surprise attack from the opponent and any attempted fall short/counter-time/draw-cut in the middle. But if your opponent is either doing the same as you, and able to take the initiative immediately and develop a fast attack, then they have advantage. Similarly, if two fencers are attacking simultaneously, it is the fencer who can reach inside the timing to find tempo, or slow the preparation to control the attack who will win. The starting point to understanding how to win the bout is an analysis of the opponent and making a series of logical tactical choices. For example, if fencers are hitting simultaneously and then a hit is scored from a fast attack, chances are that the opponent will respond with their own fast attack. If this is in to a slow preparation, to me this would be an indication of a high level of tactical control of the bout. I’d argue that the data might not necessarily show which individual preparations are most successful, but which combination of preparations that determine tactical advantage over the course of the bout.
Phil – I agree with your thesis that the person with the better control of the tactical situation, e.g. better able to adapt to the opponent’s preparation and predict their next move, will probably win. If one sabreur is clearly better than the other sabreur (e.g. in the example in this post) they will probably win regardless of what preparation-action they choose to do. That competitive edge is definitely a dimension we don’t yet consider in this set of analyses. The core assumption is that the two fencers are of comparable level and have similar reaction times and capabilities.
Once we get more data we can probably work out a way to figure out which combos of preparation-actions work better than others. For example, is the classic 1. simultaneous, 2. check fall short, 3. attack-on-preparation all it’s cracked up to be? My hunch is probably, but I don’t think I have enough data to back it up just yet.
Another element that we’re measuring right now but not really making full use of is distance. I agree with the old hypothesis that you want to keep the distance between you and your opponent in the 4m as big as possible, and let them choose to charge and get close to you. That way, you’ll always have more time to see and deal with your opponent’s choice (e.g. AoP if they run at you, start your march if they hesitate, etc). Hence the person who is better able to discipline themselves to maintain priority while maximising distance would have better odds of success. We might even have enough data to figure it out at this stage, but it’s hard figuring out from the videos who chooses to compress the distance in the middle. I’ve tried using proxies like where the back foot is after the initial advance as a proxy metric but certain preparations throw this thing totally off: e.g. it works for analysing guys like Kovalev, Montano, Homer and Szilagyi, but doesn’t work at all for any of the Koreans or Chinese and things get weird with some of the more gangly guys like Buikevich.