From poker to chess, we’re constantly putting artificial intelligence (AI) up against human opponents in increasingly complex games to see whether it can outperform humans. Yet, doing so often prompts some interesting results and discussions.
Why, for instance, can AI beat the best players in the world at Go, an ancient Chinese board game with a reputation for immense complexity, but struggle to defeat human players in Dota 2, a battle arena video game with far less intellectual credibility?
The simple answer is that while AI is exceptionally powerful, it is not suited to solving every problem (yet), with some decisions best made by humans in areas where “knowledge” is not complete. This is every bit as true in academic circles as it is when it comes to AI’s contributions to digital advertising. In the best case, there is a clear definition of success, a known playing field, and overabundance of data to learn from.
Take translation software as an example. Even though the goal “acceptable translation” is not perfectly suited for AI, we have seen tremendous success of AI-based systems in recent years. Originally, experts tried to teach rules on syntax and semantics to translation systems. These rules tried to cover all the fundamental details of language – with little success. Then software like Google Translate entered the game. Prior to this, experts tried to create a rules-based translation system with little success. Languages contain so many variables, dialects, fuzzy semantics, and definitions, it’s clear that a rule-based approach was never going to work. It needed massive data sets to learn from and solutions like Google Translate switching to deep learning – a subset of AI – to reach the level of quality where it is today. This is only possible by leveraging vast volumes of training data and fully autonomous learning at scale.
The same principle applies in digital advertising, where AI determines ideal cost-per-acquisition (CPA) based on the likelihood of conversion. There are thousands of touch points influencing the customer journey, which can be combined in countless sequences, but AI can learn from billions of signals to predict conversion probability and generate the right bid in real-time. Humans could never build business rules to achieve the same results. Let alone update and re-model their business rules frequently accounting for the range of variables and rapidly evolving conditions which define the digital ecosystem.
This exceptional talent of AI for learning in a structured environment explains why it can play Go at a level never seen before in the history of the game. Within the narrow parameters of the game, machines can analyse more possible moves than humans by combining simulations and deep learning. Just think back to the game played between Lee Sedol and AlphaGo where the computer won. It wasn’t the win that was interesting but the computer’s 37th move – one that a human would not have made – which proved pivotal later when Sedol lost. The machine’s capacity for learning, and training by playing millions of games against itself in a basement, put it not one but many steps ahead of its human opponent. These 37th move-like actions are the ones we will learn to accept, especially so for digital advertising if we want to stay at the top of our game. Sometimes, in scenarios with clear and well-defined advertising goals, the machine just knows better.
So why are machines, with all this power, still losing in the game of Dota 2? Despite the fact that this multiplayer online game seems far less intellectual than Go at first sight. It is actually more difficult for a computer to play. Players have very little information at the start, they need to explore and discover where and who their opponents are. They need more traditionally human behaviours – gut feeling and instinct to analyse the behaviour of other players, to operate in uncertain environments, recognising innate human behaviour such as bluffing and the ability to account of illogical or misleading behaviours that machines aren’t yet capable of.
While machines are making strong progress in Dota 2, it is an environment where people still currently prevail and where the value of human intelligence is exposed. In a situation where there is incomplete knowledge or a lack of data, where right and wrong is not immediately obvious, where real creativity is required, or where the space is too broad, AI takes longer to learn and human decision making is still superior. All in all, there will be a long period of coexistence, AI is amazing at certain things – but we’re still some time away from general AI that can fully act like a human.
This doesn’t mean humans should simply let the algorithms run and leave machines to get on with the tasks they are best suited to, in fact the opposite is true. Humans and machines have different skillsets and strengths. These are the most fruitful when leveraged side-by-side in a collaborative setup. People need to get better at communicating with machines, finding out what they are doing and why, to build understanding and trust in AI recommendations or automation. Only then can they let machines do the things they accel at, freeing humans up to fully leverage the innovative expertise they bring to the table. Artificial and human intelligence both have their own strengths and their place in world, and they must play the game as partners not as foes. We see this hybrid as the fundamental approach that drives the technologies we’re building and the problems we solve for clients and it brings with it a major competitive advantage for machine-enabled businesses.
The core process of identifying opportunities, evaluating data streams, and then making decisions hasn’t changed. The need to do it in real time, with an ever-increasing series of inputs from a set of channels that will only continue to grow is where machines add value and where that value is invaluable. As we play to our skills, success revolves equally on offsetting our weaknesses and that’s where partnering with machines has already re-shaped how digital advertising is executed.
Latest posts by Jochen Schlosser (see all)
- Why Humans and Machines Should be Partners Not Foes - April 25, 2019