Whoa!!
~PM
Date: Sat, 21 Dec 2013 13:43:19 -0600
Subject: Re: [agi] A Random Thought...
From: [email protected]
To: [email protected]
Ok, which dimension are you attempting to scale up? Triviality corresponds to
minimal representational capacity, both in the environment and the agent
operating within it. Human beings are (currently) at the other end of that
scale, with enormous representational capacity for dealing with a highly
complex environment. The more complex (non-trivial) the environment, the
greater the representational capacity required of agents operating within it in
order to effectively make decisions. It is this dimension that I am looking at.
Learning algorithms are easy to understand, design, and implement. They are
just solutions to optimization problems. I do not think learning itself is
where the bottleneck lies. Instead I look at the representational systems
underlying those learning algorithms. The simplest learning algorithms operate
over tables of choices. They tabulate expected returns or error levels for each
choice, over many repetitions, and gradually settle on the choice(s) with the
maximum expected return or minimum expected error level. Adding layers of
sophistication, we begin to see context matter more and more: Conditional
choices and statefulness result in much more interesting and coherent behavior.
Generalizing over choices and conditions and actions to those that are similar,
we see an additional gain in coherency, with algorithms that can deal with new
situations robustly based on previous experience with other situations.
What is needed is to increase the expressivity of the underlying
representational schemes used by learning algorithms. Moving up to the
representational complexity level of ontologies, episodic memory, etc., the
representational scheme becomes ever more capable. In order to reason about
things, we need to represent those things effectively. Once we have a fully
capable representational scheme -- a programmatic framework for the
representation of Meaning, in all its forms, with all its inherent ambiguities
-- we can begin writing learning algorithms to extract meaning from the
environment, generate rules for predicting arbitrary unobserved phenomena from
arbitrary observed phenomena, recombining meanings to produce new ones,
choosing contextually appropriate and meaningful behavior, etc. There is no
understanding without meaning, and there is no intelligence without
understanding.
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
