On 1/19/07, Benjamin Goertzel <[EMAIL PROTECTED]> wrote:
You have not explained how you will overcome the issues that plagued GOFAI, such as -- the need for massive amounts of highly uncertain background knowledge to make real-world commonsense inferences
Precisely, we need to amass millions of pieces of knowledge items, and some items may have uncertainty. This is precisely what I'm trying to do. The alternative route is machine learning, but that requires a sensorium or a NL interface, which is an even more daunting task. (But I don't object you going that way =))
-- the combinatorial explosion that ensues when you try to control logical inference on a large body of data
I have studied this issue a bit, unbeknown to you =) There are ways to tackle massive numbers of rules, eg the rete algorithm, predicate hashing, etc. Soar is a good example using the rete algorithm. It can handle millions of rules (and probably many more).
My own solution to these problems is to -- learn most knowledge via experience rather than via explicit encoding
Nothing wrong with this approach, but it may be even more difficult than mine.
-- utilize a subtle combination of inference, statistical pattern mining and artificial economics for inference control
You're getting into the topic of inference control, but I was only talking about collecting knowledge in the form of rules. Speaking of my project, it does not endorse specific inference methods. It is up to the AGI designer how to use the data. BTW, statistical pattern mining is good for *learning* patterns, I wouldn't use it for inference per se. For me inference is done only using *existing* rules and facts in the KB. Pattern mining is for discovering *new* rules and facts, which is very time-consuming and compute-intensive.
Pei agrees with me on the "learning via experience" part but has a different approach to the combinatorial explosion problem of inference control. But you have not yet presented any original solutions to these or other major well-documented problems with the GOFAI approach.
Again, I have no problems with "learning via experience". What I propose is to augment this with "knowledge acquisition via direct encoding, with the help of the net community". Do you have some reasons against this? Is it difficult for Novamente to incorporate the rules database?
Yes, intuitively the approach you're suggesting sounds like it should work -- at first. That is why masses of research funding were spent on it decades ago, and why hundreds of brilliant people spent their lives on GOFAI. But you are not giving us any rational reason to suspect you might succeed in this sort of approach where so many others have failed. What is your new and different idea?
I think the key innovation is that I allow rules with variables as well as facts, and that such knowledge would be collected from online users on a massive scale (which doesn't mean the project require massive $$$s). Such a combination has NOT been attempted before, AFAIK. Frankly I'm not that knowledgeable about failed GOFAI projects (I was just teenage in the 80s, playing with a TRS-80). Decades ago, the internet didn't exist and there was no way of amassing knowledge like MindPixel can. This is perhaps the most important reason why past projects failed. Maybe you're just reflexively saying that GOFAI is a failure, without giving it serious consideration. Cyc is not a complete failure. It's still ALIVE and it can do some reasoning about terrorist attacks etc. Why wouldn't an improved GOFAI succeed? Perhaps it is an misconception that everything associated with GOFAI _must_ be abandoned... YKY ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
