If you haven't seen the classic documentary King of Kong, do yourself a favour and watch it. It delves into the bizarre and hilarious world of professional arcade gamers. It shows the immense level of dedication and obsession involved in mastering an old arcade game. It's impressive then that the AI of today is able to easily beat the best in the world at many of these arcade games.
Recently there have been quite a few headlines about AI computers beating world champions at various games like chess, DoTA and GO, but it's hard to see how this has any real significance. What does this mean for the world today?
The answer is: quite a lot.
Think about it. If an AI that knows nothing about a game can, in a matter of hours, learn how to play it as well as a human that has practiced for many years, then by taking real world tasks and framing them like a 'game', the same AI can learn to do these tasks at a superhuman level.
Take the example of pacman, a favourite in the fish and chip shops of the 80's. There are many real world problems that can be framed like pacman. First lets consider the key elements of the task:
- The goal is to get a top score by moving pacman around the screen
- Pacman must collect various items in the most efficient way possible while avoiding ghosts
- At times the rules change and pacman can also eat ghosts for more points
The best players learn advanced strategies and routes that maximise their score. AI's have been observed to emulate these strategies.
Consider a second 'game': GarbageMan. The rules are simple:
- The aim is to achieve the best score by moving GarbageMan around and collecting garbage from bins.
- The longer a bin is full, the lower the score when it is collected.
- When the truck is full, it must return to the tip to be emptied.
Sounds like fun right?
Imagine if you had someone with the dedication of a pro arcade gamer spend hours every day playing GarbageMan. He would learn every route, the rates at which bins become full and strategise about when to empty his truck.
In case you hadn't realised, GarbageMan is not a real game. It's an optimisation problem faced by many cities around the world. The fact that it can be framed like a game is what's so exciting, because the same AI player that mastered pacman in minutes can also master GarbageMan.
That's one example, here's a few more 'games':
- AmbulanceConrolCentreMan
- ParkingInspectorMan
- SupplyChainPlannerMan
You get the idea...
The field of teaching AIs to play 'games' is called reinforcement learning. It's one of the most exciting areas of applied AI and one of our main focuses here at Piccard.
If you're interested in finding out how AI can achieve outstanding results in your industry, please get in touch.