How does Marketing Intelligence affect Computer Learning?
To win in worlds of infinite options, you need the best of both: human intelligence to limit the first cut of possibilities, and a self-learning machine algorithm which can then execute, compute and analyze far better than any human
Thirty years ago, I played in an international chess tournament in Eastburn, England. In the 4th round, I found myself sitting in front of a computer. It was the first and last time I played against a computer in a tournament. I won the match in a surprising way: the computer performed a “blunder” move which lost his rook to my bishop. It surprised me because I thought I would have to win the match using my intelligence and strategic understanding of chess. However, I ended up using my rudimentary tactics: the bread and butter of computers. Of course, let’s not forget it was a “weak” computer, and times have changed a lot since.
Computer chess is very different from professional chess. While the computer can calculate millions of moves per second using brute force, humans only calculate the most “logical” moves. The biggest advantage for professional chess players is their ability to understand the strategic aspects of the game, the best positions on the board, and the irrelevant moves.
The Limitations of Brute Force
The Brute Force approach is best at solving problems where the number of options is limited. A good example is a Rubik’s Cube: even when you think you saw the best human attempt at it, the machine will still beat him.
However, in chess, the number of possible moves is infinite: there are 400 different options after each player makes one move apiece. 72,084 positions after each player makes two moves apiece. More than 9 million unique positions after the third move. After the 4th move, more than 288 million different positions are possible. There are more game-trees of chess than the number of galaxies (100+ billion), and more openings, defenses, gambits, etc. than the number of quarks in the universe!
So, in the case of chess, brute force won’t win the game even for a monster computer that can calculate hundreds of millions of moves per second.
A Layer of Intelligence
All of this changed about 25 years ago. On February 10th 1996, for the first time, a computer won a game of chess against a world champion. The computer was IBM’s Deep Blue and the world champion was Garry Kasparov, considered by many (myself included) to be the best chess player ever. This was the first game in a best of six match, which Kasparov finally won 4:2.
After the match, the IBM engineers asked for a rematch. Leading to the game, they focused on the artificial intelligence aspects of their software: “teaching the computer to think” by feeding the computer games that Kasparov played before. Famous chess Grand Masters on the Deep Blue team helped the engineers develop the intelligence and understand the game: a substitute for calculating all possible moves.
The rematch was set for May, 1997. Kasparov won the first game by playing strategically based on “anti-computer” tactics. The turning point came during the 2nd game of the match. The computer won this game by playing an “intelligent” game. In one of its moves, it even preferred not to win an immediate “benefit” of taking a pawn, which Kasparov thought would give him some strategic advantage. Instead, it played a move that improved its strategic position.
The match continued with 3 draws, but in the final game the computer beat Kasparov who was still frustrated from the 2nd game – and finally won the match. After the match, Kasparov accused IBM of cheating because he couldn’t understand how the computer was able to think that way.
Technological progress has improved chess engines considerably. The quantum leaps were not in the realm of computational force, but rather in the “intelligent” factors of the software. Just like a human player, an engine doesn’t look at all moves to the same depth. Potentially good moves are examined exhaustively, whereas weaker moves are only given a quick, rudimentary look. It’s similar to the instinct and experience of a strong human chess player – looking at just a handful of moves in a position, discarding the others nearly instantaneously.
Combination of Marketing Intelligence and Computer Learning
In my youth, I was able to beat the most rudimentary chess computer, though the more “intelligent” versions which came out in subsequent years left me far behind. Luckily, my interests shifted from chess to software. Today, I can implement the logic from the chess software that beat me and create marketing technology that beats the competition. Here is what we know:
- There are endless customer personas. Every customer can have hundreds of different attributes. There is an endless amount of offers available in the e-commerce world: any kind of discount on any kind of item, free shipping, 1+1 etc.
- There are many ways to measure the success of every campaign: impressions, orders, order days, time spent on site, etc. There are also many communication channels with the customer – email, SMS, push notifications, social channels and so on. The approach of testing every possible incentive for every customer attribute in each communication channel is very similar to the old computer approach to chess, which examined every legal move on the board, putting 99% of its computing power to waste. If we were to use a naïve algorithm of brute force in a marketing scenario, the combinatory amount of different possibilities will cause the computer to take forever and end up with no conclusion or loss of customers.
- There are many possibilities that a marketer wouldn’t even consider. For example: sending a European football fan an incentive on a 4th division Australian soccer game on the day of the World Cup finals, or sending a 2% discount during Black Friday. While the computer will have to test these options in order to understand whether they work or not, it would not even cross the typical marketer’s mind.
So, brute computational force will not lend us an advantage in marketing. I believe that the way to win is for the marketer to instruct the machine based on a target group of customers: this will show their lifecycle stage, specific attributes, activity history, etc. The marketer will also limit the frequency, incentives and channels – and the machine will take it from there.
Machine Learning Magic
This is the precise recipe for Optibot – Optimove’s marketing optimization bot. After the marketer defines the target groups and a number of possible offers, it’s Optibot’s job to do the magic. It runs the offers for different groups in the segment, measures success rates and re-divides the groups accordingly, until one incentive beats the rest, meaning that we have found the statistically significant best action for this group. We call this a “self-optimizing campaign”.
During the process, Optibot can decide how to further divide the segment to a few smaller sub-segments (for example, by using four different Customer Activity (RFM) levels: very high past activity, many high past orders, few mid past orders, and several low past orders.) Based on these campaigns we can find which action will work better for each sub-segment.
Of course, this is something that will be nearly impossible for a marketer to do because there are just too many small decisions to make, countless variations of how to divide the proportion of the incentives for each group, and an endless number of frequencies for running the campaign.
However, it’s a walk in the park for Optibot, which utilizes machine learning algorithms and self-learning mechanisms to enable it to predict the success rate for any incentive, learn from the real results and tune the algorithm accordingly. This is just what advanced chess programs do when they evaluate the position before the move, rate each possible response, and then examine if the prediction of the counter move was correct and if the evaluation of the new position changed.
In order to win in worlds of infinite options, you need the best of both: human intelligence to limit the first cut of possibilities and create a “closed environment” for the machine, and a self-learning machine algorithm which can then execute, compute and analyze far better than any human.