Good points. ~PM Date: Mon, 28 Apr 2014 19:29:09 -0500 Subject: Re: [agi] Situations From: [email protected] To: [email protected]
A situation can be modeled as a set of related entities together with a set of constraints. Think of it as a partially determined simulation or description whose scope can be expanded as reasoning or analysis proceeds. (Note that this expansion of scope need not be restricted to a particular temporal direction.) Another way to look at it is as a collection of phenomenal binding sites (like noun phrases, or fields in the SELECT portion of a SQL query) and relationships between those binding sites (analogous to verb clauses & prepositional relations, or conditions in the WHERE portion of a SQL query). As such, a situation can be seen as a compound pattern or rule, which can potentially be matched against sequences of states of the environment. One situation can be considered a generalization of another if it can match all of the environmental state sequences that can be matched by the other, while preserving the mapping from entities to components of the environmental state sequence that allowed the other situation to match. The most general or natural way to represent a situation, I think, would be as a graph in which vertices are labeled with descriptions of entities and edges are labeled with relationships those entities must abide by. Use of the term "irrelevant" towards situations assumes a purpose or goal to which a situation may be relevant. To determine whether a situation is relevant with respect to a particular goal, several conditions must be met by the situation. It must be possible to reach the situation from the current situation with non-trivial probability and tolerable cost, and the perceived distance or cost to the goal must be reduced or maintained. The A* algorithm comes to mind, but in this case it would be operating on a meta-graph -- with situation graphs as vertices -- whose topology is dynamically generated by applying transformational rules to generate candidate future situations from current ones. For example, supposing the goal is to enjoy an apple, and the situation is that you have an apple which is dirty, three of the many transformations which could be applied are: Eat the apple immediately. This transformation results in a relevant situation due to the reduced distance to goal, despite increased cost from reduced enjoyment.Wash the apple. This transformation results in a relevant situation due to the reduced cost to the goal, despite increased goal distance due to the delay. Throw the apple at the wall. This transformation neither reduces cost nor distance to the goal, and thus produces an irrelevant situation.In each case, the new situation must be constructed by applying a graph transformation associated with the proposed action before it can be determined if it is truly relevant. By recording the situation/action/success triples from previous experience, and searching for such triples that maximize the intersection (i.e. similarity) between initial situations and the one that we are currently facing, it should be possible to heuristically estimate the utility of the given action within the current situation, allowing us to sort actions according to likely utility and thereby avoid processing of irrelevant actions and situations until better options have already been exhausted. On Mon, Apr 28, 2014 at 1:35 AM, Piaget Modeler <[email protected]> wrote: How do we form situations in our mind? Some may be actual, hypothetical, or anticipatory. How would you model situations? Assuming that we have millions of them to choose from, how do we ignore irrelevant situations and work with relevant ones? I have some theories, but I'd like to hear your thoughts? ~PM AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription ------------------------------------------- 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
