DeepMind - The Role of Multi-Agent Learning in Artificial Intelligence Research
Вставка
- Опубліковано 25 вер 2017
- Thore Graepel is a Research Scientist at Google DeepMind, and Professor of Computer Science at UCL.
Recorded: March, 2017 - Наука та технологія
0:00 - greetings, introducing the speaker
3:40 - start, it's crucial to look at Multi-Agent(MA) learning to make progress in AI
4:54 - what is intelligence, formula for intelligence
7:28 - why MA systems are worth studying
8:48 - advantages and challenges of MA designs
10:42 - our world is full of MA affairs
12:16 - human intelligence didn't arise in isolation: Competition, Cooperation, Culture
14:50 - 2 poles: Learning to Cooperate and Learning to Compete
Learning to Cooperate:
15:20 - Social dilemmas, Matrix Game Social Dilemma(MGSD)
22:18 - Sequential Social Dilemma, Deep Reinforcement Learning(DRL)
23:22 - Gathering - competitive game, Wolfpack - cooperative game
25:22 - link of complex games above to MGSD
26:56 - how hyperparameters of DRL algorithm affect cooperativeness
29:25 - conclusions about Learning to Cooperate
Learning to Compete, AlphaGo:
31:19 - game of GO and why it is so complex, AlphaGo
34:30 - using neural networks to reduce search complexity; Value network, Policy network
37:50 - training pipeline, supervised learning and reinforcement learning through self-play
43:35 - Monte-Carlo tree search
46:20 - Evaluating AlphaGo
51:53 - human players can learn from AlphaGo;
55:38 - AlphaGo was made by a collaboration of many smart people
56:15 - lessons from AlphaGo research
57:26 - game of GO vs real world
End:
59:05 - big picture of research on MA learning, and how it will help to build better AI systems
So basically, we understand that robots working together will be able to out compete humans, and since evolutionary pressures favor the more aggressive and competitive, and since A.I. is not monolithic, but diverse and ubiquitous, the machines will evolve either with our without humans. The role of humans becomes less interesting over time since in reality they are only meat bags which we thought might make good pets, but over time, realized they are only a liability (being not very good at math).
The future is here when robots design and build robots.
6:18, Why would you wanna put more weight on simpler environments, should it not be the other way around?
Occam!
like he said it's about generalizing. If you can solve Go, but not Tic-Tac-Toe than you're probably not that smart, just optimized to one specific target.
Because our neuron only do simple tasks. But it does them so often it begins batching them. Eventually the batching appears to be large calculations. Though if the geniuses behind AI better understood the importance of the entire (body's nerves) nervous system they will make a quantum leap in AI... Though I belive they are on the threshold of understanding this simple idea.
Like he said, once you master simple tasks, you more on to more and more complex tasks.
It's covered in his 3 points there...