like i said it's reasonable to drive for all at different points in the past but it's unreasonable to drive for all 4 at the same time. Think about it, incentive-wise, if you complete 50 trips you got $x , if you complete 100 trips you got $2x. If you only got 50 trips in you, why are you splitting them between two apps 25 trips each and got $0 incentive?
it's economically unreasonable to drive for more than one apps simultaneously. Now, it could be economically reasonable to drive for more than one apps at many points in the past.
If Lyft and Uber are so easily exchangeable, why is Lyft is still a minority in the US while spending more money?
> Uber has a global operations, is in multiple streams of business and has diversity across business lines
This can be a reason why one would be more interested in Lyft. Uber seems like a distracted player who's losing money on many other markets and businesses, not to mention hundreds of millions of dollars on self-driving cars (and flying cars?!). Lyft is much cheaper (15B valuation), while Uber is much more expensive (120B?). If I invest 1B in Uber, my money would vanish in 1 quarter (yes they're losing 1B/quarter). Those 1B dollars would be split to invest in flying cars, uber eats freight bike/scooter, battles in India Middle East. On the other hand, if I invest 1B in Lyft, I'm sure those 1B would go towards gaining market shares in the US which is by far the most important market for the two players.
Second of all, personally I think if Lyft failed and the stock dropped by half. Some other dominant players would look to acquire Lyft. I'm thinking about Google's Waymo One plus Lyft's network. Apple seems to have a lot of cash to burn also, and they're also developing SDC. On the other hand, Uber's share price has to drop more than 10x in order for it to come close to a reasonable acquisition price.
There's a long version and a short version. The long version is you have to learn to write mathematics by yourself. Start with an intro course and start deriving theorems by yourself. Do not look at the proofs. At this stage, details are very important and can't be overlooked. You need to be your own critic and keep asking why and how to every single detail and step until you can convince yourself that you would be able to naturally come up with the theorem and proof. Continue doing this to higher level courses. This is how I learned Math since middle school all the way throughout graduate school.
The short version is you have to ask the right questions. Naturally for every theorem or equation, there are 3 big questions:
1) What does the theorem/equation say? What's the intuition behind it?
2) Why is it true?
3) How does one come up with it?
One must ask these questions in the exact order. To understand what the equation really means, you should break it down further to smaller components. What is this variable? What does it represent? What is the intuition behind what it represents? What's the implication when the variable increases, decreases, etc? Do that for every single component in the equation/theorem. One should fully understand the intuition and clearly describe all quantities before trying to look at the equation/theorem as a whole.
To understand why an equation/theorem is true you need to build up a repertoire of theorems related to the quantities of interest. The bigger your repertoire, the easier you can prove or disprove something. The more advanced way is to build up intuition around the quantities of interest then come up with intuitive hypotheses. The hypotheses are often easier to prove/disprove. The process repeats.
- model free methods have seen great success in terms of learning high dimensional tasks however it suffers from being sample inefficient. In other words, it takes too long for real robots. Examples of these methods are TRPO, PPO, ES, etc
- model based methods is an order of magnitude more efficient, and thus, are more practical on real world robots. However, these methods have high bias and most working models are simple in terms of representation power, e.g. GP, time varying linear, mixture of Gaussians,. Examples are PILCO, GPS, PETS, etc
Of course, SOTA is a lot more complicated but it's a short explanation to your observation.
From the description it does sound like blink dagger and the range here refers to radiance or necro's heartstopper. It's definitely not "previously undiscovered". Also, the article makes it sound like we saw another AlphaGo's 3-3 invasion, 5th line shoulder-hit kind of moment. We did not. This is more similar to AlphaGo and Fan Hui match, except imagine AlphaGo lost to Fan Hui. The bots did make a lot of interesting moves in 5 invincible chicken meta. The bots appeared very weak in normal meta (constantly check rosh for no reason, inefficient use of ults, don't get me started on warding, etc)
The surprising thing to note here is how much optimism the street has in $twtr. Just one quarter of small beat and the stock jumped 100%. After the drop today (to $34) it's still too high for $twtr which hasn't actually proved anything since the stock was high teens low 20s. If anything, their live streaming effort is pretty much down the drain, and Anthony Noto, who's pretty much the heart and soul of Twitter operation, left. I'm super surprised the stock is still mid 30s.
Very impressed with some of Voyage's recent hires. A question for @olivercameron,
What makes Voyage different? From what I understand, you pick canonical routes inside private communities. Let's assume demand on these routes are high enough, and there are enough private communities to make a significant market. What prevents Google from coming in and mapping the area in a week and run you out of business? Let's say, hypothetically, I'm a self driving car engineer, why would I pick Voyage over other big players who have a lot more capital and much bigger team with a lot more people like Drew Gray?
Yes but no one plays serious in All Random mode. Supports are called supports for a reason, they're strong early game without a lot of items. Some gave early ganking, counter-ganking abilities, some have healing, harrasing abilities or really good early stats. Carries are stronger mid game and late game because naturally they're weaker early game. Because of this, you need to specify farming position, setup ganking, rotation, early game. All of these strategies will be gone if you play All Random. Basically it becomes a 2k pub trash game and no one > 4k mmr actually practices all random daily. Unless you're in SEA where people just first pick carries :) :) :)
Does anyone know how random drafting work? If they're truly random, i.e. randomly picking 5 heroes out of 18 (CM, DP, ES, Gyro, Lich, Lion, Necro, Qop, Razor, Riki, Nevermore, Slark, Sniper, Sven, Tide, Viper, or WD) then it's much less about teamwork. What if they end up with Razor, QOP, Nevermore, DP, Gyro? The problem here is in Dota, each hero almost has a clear position in the game, much like soccer. Having both teams randomly pick 5 heroes would probably ruin the game, and make it really difficult for human (think covariate shift), whereas the bot is probably trained using this distribution
I'm a Legend dota2 player and also a Machine Learning researcher and I'm fascinated by this result. The main message I take away is, we might already have powerful enough methods (in terms of learning capabilities), and we're limited by hardware (this also makes me a little sad). My thoughts,
1) "At the beginning of each training game, we randomly "assign" each hero to some subset of lanes and penalize it for straying from those lanes until a randomly-chosen time in the game...." Combining this with "team spirit" (weighted combined reward - networth, k/d/a). They were able to learn early game movement for position 4 (farming priority position). For roaming position, identifying which lane to start out with, what timing should I leave the lane to have the biggest impact, how should I gank other lanes are very difficult. I'm very surprised that very complex reasoning can be learned from this simple setup.
2) Sacrificing safe-lane to control enemy's jungle requires overcoming local minimum (considering the rewards), and successfully assign credits over a very very long horizon. I'm very surprised they were able to achieve this with PPO + LSTM. However, one asterik here is if we look at the draft, Sniper, Lich, CM, Viper, Necro. This draft is very versatile with Viper and Necro can play any lane. This draft is also very strong in laning phase and mid game. Whoever win sniper's lane and win laning phase in general is probably going to win. So this makes it a little bit less of a local optimal. (In contrast to having some safe lane heroes that require a lot of farm).
3) "Deviated from current playstyle in a few areas, such as giving support heroes (which usually do not take priority for resources) lots of early experience and gold." Support heroes are strong early game and doesn't require a lot items to be useful in combat. Especially with this draft, CM with enough exp (or a blink, or good positioning) can solo kill almost any hero. So it's not too surprising if CM takes some farm early game, especially when Viper and Necro are naturally strong and doesn't need too much of farm (they still do, but not as much as sniper). This observation is quite interesting, but maybe not something completely new as it might sound like.
4) "Pushed the transitions from early- to mid-game faster than its opponents. It did this by: (1) setting up successful ganks (when players move around the map to ambush an enemy hero — see animation) when players overextended in their lane, and (2) by grouping up to take towers before the opponents could organize a counterplay." I'm a little bit skeptical of this observation. I think with this draft, whoever wins the laning phase will be able to take next objectives much faster. And winning the laning phase is really 1v1 skill since both Lich and CM are not really roaming heroes. If you just look at their winning games and draw conclusion, it will be biased.
5) This draft is also very low mobility. All 5 heroes Sniper, Lich, CM, Necro, Viper share the weakness of small movement speed (except for maybe Lich). Also, none of these heroes can go at Sniper in mid/late game, so if you have better positioning + reaction time, you'll probably win.
Overall, I think this is a great step and great achievement (with some caveats I noted above). As far as next steps, I would love to see if they can try meta-learned agent where they don't have to train from scratch for a new draft. I would love to see they learn item building, courier usage instead of using scripts. I would also love to see they learn drafting (can be simply phrased as a supervised problem). I'm pretty excited about this project, hopefully they release a white paper with some more details so we can try to replicate.
If you have money, you have domain expertise. They hired Mark Moore from NASA and Celina Mikolajczak from Tesla. Each of these people can hire an entire team of experts from NASA and Tesla.
It's not about needing it in "the real world". It's about having fun. It's about the adrenalin rushing through your veins as the clock ticking down. It's about burst of joy when your brain just clicks and the invisible wall falls down, showing a path to the solution. I still do Putnam and IMO every year, I also occasionally take an hour or two during work and just solve Math problems. Solving problems is a hobby, just like video games. There's nothin really practical about it and there doesn't have to be.
I don't understand the surprises here. New grads out of undergrad these days are making 120k in base + about 250k RSUs vested in 4 years + 20k cash bonus every year. That's about 200k / year for a new grad from UNDERgrad. Ian Goodfellow invented GAN and he's paid 1M a year and people are shocked?