PF News

Oh Psi! All that work! What a shame! I at least hope the participants understand the significance of crippling a bot right before the contest. Like you said, you made good contacts. If they take the time to try BJ when all is working, they will be impressed.

Good show in getting that far! I am sure you learned a lot and saw some interesting things. Remember no one "wins" the Loebner unless they effectively create a Data from Start Trek or a less insane HAL. BJ rocks, and so do you!

I vote we work on sending the evil Hal..Muuhhaaaaaa!

I have noticed that 'luck' seems to play as big a part in contests as the quality of the bot.

That is very true - tests composed entirely or predominantly of trick questions and fatuous trivia test nothing more than the patience of the botmaster. Still, the fatuous trivia at least can be dealt with with an external ontological knowledgebase, so that's probably next on the agenda.

Oh Psi! All that work! What a shame! I at least hope the participants understand the significance of crippling a bot right before the contest.

Yes, they're all nice guys, and everyone was sympathetic, but what can you do.
It was great to finally meet people I've sort of known second hand for so long, and to put faces to the names.
And I at least have the consolation that BJ's beaten them all 2 years out of 3 in the CBC (and thanks to Bildgesmythe, the Forge has thrashed them 3 years out of 3! ) Even allowing a fair portion of luck, we are demonstrably using the best bot development platform there is, bar none!

True, but we'd have to be able to combine the PF type engine with a true learning AI to get to the next level. I'll just be here holding my breathe.

PS I think I like "Psim" better than "Psi" now--makes me think of Psimulating personalities (or at least your psim in SL).

I don't think we have exhausted the ability of Forge bots to learn through memory. Since a memory can be created from any information that a conversation provides, Forge bots are capable of learning a great deal.

How would you make it remember an entire sentence? Let's say the guest says something, and you want to remember the whole thing.

My theory is that you would have to create a plug-in that contains a number of frequently used words (in, or, of, about, ....). Let's call the plug-in (fuw).

As a prerequisite you need to devise a key phrase / seek system leading the guest to where the plug-in (fuw) will be triggered as a seek.

Now it is quite easy to store the response in three separate memories:
- rem (prekey) as "pre-fuw"; rem (key1) as "FUW"; rem (postkey) as "post-fuw";

Add a (*) seek to the (fuw) seek as a catch all key phrase to construct a complex memory like:

- rem ((firstprekey)(firstkey1)(firstpostkey)) as "questsaid";

I think it should work, but, it's only a theory and I am not sure the engine is happy with the complex memory construct.

LarsB that's a good idea I will have to give it a try. I haven't been as interested in "the whole sentence" as in the properties, concepts and conditions the sentence would relate. Things such as the quests opinion of where they live, the qualities of their family and friends, political views etc. I wonder how often we, as humans, can repeat the previous sentence accurately.

Okay, my theory is useless, perhaps the Prof can implement a solution where the following errors will be resolved, i.e. where it will be possible to indeed store a constructed memory. Maybe the issue has been raised before, it may well be impossible to implement such a solution and I am the only one not knowing it:

ERROR: The following AIScript statement is incorrectly formatted: "rem ((firstprekey)(firstkey1)(firstpostkey)) as "guestsaid"". Please correct this and try again.

And also:

ERROR: The following AIScript statement is incorrectly formatted: "rem ((prekey)(key1)(postkey)) as "guestsaid"". Please correct this and try again.

I'd like to read other approaches because it is a very interesting capability for a bot to have.

Prob, I don't think from memory has been exhausted and I think the Forge bots are wonderful, but at the end of the day, it still rests on a person thinking up almost every possible KP and having several responses, even if they are personalized with memory. As they are, the PF bots will never get context through memory or apply memories in new situations. If I tell Dragon breath "The sky is blue" you many have him say--"Hey didn't you tell me the sky is blue." but if tell letter tell him "The sky is green" he won't, on his own, say "You told me the sky is blue. I have checked your assertion against other records and it seems that most sources report the sky is blue." That may not be all I want from learning and context, but can you see what I mean?

True, the bot wouldn't be able to do statistics, but
the sky is (*). Remember <?PF rem (key1) as "sky" ?>
The sky is blue will be remembered, the sky is big, or green or smoggy. The bot now has a bunch of sky like properties. In the xmemory or xnone area you can have responses that will pull up memories that the bot has accumulated on any memory. "I have heard it said that the sky is (mem-sky). or a separate keyphrase for the bots knowledge base of what the sky is.
the sky is (*). Remember <?PF rem (key1) as "sky" ?>
I have heard that the sky is (mem-sky)

or better to but the memories under self mem
<?PF self: remember (key1) as only "sky"; ?> So the memories would not be defendant on the person being chatted with.
The 20 question game could be reproduced. Osiris already does something similar with personal info.