Ask not if AlphaZero can beat humans in Go — Ask if AlphaZero can teach humans to be a Go champion !!!
1. On Intelligence & the Art of a (ø)Go Game Engine
In many ways, AlphaZero is an advanced toaster or a calculator. A calculator, however advanced it is, can’t teach algebra nor a toaster deliver tasty meals. Reason being, they have neither the semantic understanding of their function nor the benefit of human knowledge. Even Watson is oblivious of the human common-sense. I had written about this question before w.r.t IBM Watson during the Jeopardy match (Section 3 of that blog).
A recent example is Alexa sending a random piece of living room conversation to an employee.
If Alexa had the context and common-sense, it would have definitely concluded that random string of commands garnished from conversation snippets do not form a cohesive action.
IMHO, Alexa has an identity crisis — it is mostly a Q & A machine — but it thinks that it is a Dialogue Machine or even a Conversation Machine ! And so it tries to behave like one … with disastrous corner cases …
[Update June 2020] A more recent diagram about Conversational AI for the Transformers tutorial at GTC 2020 Digital …
2. The education of a machine-(ø)Go
Anyway, the whole musing is a prologue to something I am working on — a Semantic Go Game Engine (shamelessly dubbed øGo) that can comprehend what it is doing and thus can teach ! The ultimate goal is to develop Socially Assistive Robots that can augment humans …
At a technology level, øGo employs Reinforcement Learning, Bootstrap Learning, MCTS and so forth — a topic for a series of future blogs; here let me talk about the books !
Am putting øGo on a diet of Go books — from the game basics to openings to placements to proverbs to an overdose of Lee Sedol games !
- øGo should learn Go and have the personality of Lee Sedol in it’s thinking !
- øGo should be able to evaluate the Elo of a human playing with it & adjust accordingly.
- Plus øGo should be able to annotate, guide the human player at their level & help them improve
- Which means, øGo should not only be able to traverse a knowledge graph, but also able to overlay the player’s graph and know the knowledge disparity …
The Go set is modest but original Yellow Mountain etched bamboo board and Double Convex Yunzi Stones /Bamboo Bowls.
I have a set of basic books to start the learnin
The more interesting books are the game books — I focus on Lee Sedol games. Was fortunate enough to find some good books.
- Relentless is a huge 600 page book ! 48 games between Lee Sedol and Gu Li, hard cover and commented ! I hope øGo learns these games well.
- The Celestial Arsenal is another interesting book
- The three volume Commented Games by Lee Sedol is one of the key pieces — am not sure yet how to capture the intuition and mindset of Lee Sedol into øGo. Is bootstrap learning the right method ?
- The Theory & Practice of Go has a rich history — published in 1880, by a German Go enthusiast!
- Two books by Iwamoto Karou, Invasion and Reduction are classic Go
- I hope øGo listens to the 3 Volume Lectures in Go techniques and then refines it’s techniques with the 2 Volume 21st Century New Openings !
- And finally have Joseki debates between different versions of øGo & take the proverbs to it’s positronic heart ! (Incorporating the proverbs into the behavior of øGo would be an interesting challenge — somewhat akin to Asimov’s 3 rules!)
As I said earlier, from an algorithmic perspective, we have the Knowledge Graph, MCTS plus the various Reinforcement Learning techniques. Are they enough to really develop an interactive Go teacher ? I have no clue ;o( … And, definitely an year+ long project …
