Alpha Zero is the game playing AI.

One might try to use the algorithm in a robot which takes input from it's environment and thinks about taking the best course of action.

So I was thinking about what is missing from Alpha Zero that would not make this work.

One thing, I think, is that Alpha Zero is pre-programmed with the rules of the game it is playing. Wheres I think a general AI should learn what effect it's actions have. This could be done with a neural network to simulate the next step in a game given various actions.

Also I think it needs some sort of attention, since in Chess there are only a finite number of moves, but a human has potentially many more. But a move such as "pick up apple" is not good if there is no apple around. But "pick up thing in my attention" would work better.

I think it would need some sort of RNN module to generate or recieve sentences.

If it generated say 5 sentences, (related to what it has in it's attention) it could see these as "moves" and decided to either play one of these moves or not. This is somewhat similar to what a human does when they think about what to say in an awkward situation.

Anything else?


One reason Alpha Zero doesn't have human-level intelligence is the lack of computing power in modern day computers when compared to the human brain. The greater level of computing power allows the brain to perform more complex calculations. Less computing power translates to fewer calculations which in turn translates to less conditional probabilities (which is effectively what a neural network calculates) which means less intelligent behaviour (as the neural network can handle fewer samples).

Another possible reason is that the algorithms in the brain are superior to those of modern neural networks. For instance, humans can easily identify objects based only one or two previous encounters while neural networks have to be trained with huge volumes of data before they work.

Human behaviour might also be non-algorithmic, as Roger Penrose argues. The example he gives (in "The Emperor's New Mind") is that of a computer trying to solve Turing's halting problem. The computer can be in one of two states – the first one in which the Turing machine halts and the second one in which the Turing machine does not halt. The system is deterministic (every event is caused by a preceding one) but cannot be algorithmically predicted. Penrose argues that the human brain works a little like this, through a quantum wave collapse. This stands in contrast to beliefs that the behaviour of the brain is fundamentally algorithmic; if this was the case more and more computing power as well as better algorithms would eventually lead to an AI on par with humans.

He also believes that it is impossible to describe the human brain given the state of today's physics – a theory of quantum gravity (describing gravity at small scales and the gravity of elementary particles) would be needed to have a shot at doing so.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.