Less speculatively, one of the questions that interests me (and Andrew) is whether a Minstrel-style system might be useful in an interactive context.

I’m fairly naive about IF, but I did speculate a bit about running Minstrel “backwards” to go from episodes to higher-level structures. In a very simplistic way this is how Minstrel begins a story — by getting a bit of input the reminds it of a specific memory, which in turn reminds it of a plot. Something of the same sort could be attempted for interactive fiction. Supposing that the human in the loop enters some action, Minstrel could then try to be reminded “backwards” of some plot structure that action could fit. For example, the player drinks a potion; that reminds Minstrel of “Romeo & Juliet.” Then Minstrel could try to inject elements into the story to complete a Romeo & Juliet-like plot. Creativity would be useful both in recognizing a past plot and creating the new plot elements

Right now, the best way I can think of to create an interactive media experience around something like Minstrel is to offload the commonsense reasoning on the audience.

I had to be careful in creating Minstrel not to inject my own intellect into the results, but it seems clear to me that I could have improved Minstrel’s output considerably just with better use of language and “fleshing out” Minstrel’s reasoning. So the notion of a collaborative system which uses human reasoning to augment the computer’s creativity is intriguing. On the other hand, you’d probably get people arguing that creativity is exactly the part that people do best!

Less speculatively, one of the questions that interests me (and Andrew) is whether a Minstrel-style system might be useful in an interactive context. I think interaction is one of the ways that audiences can really understand the power of generative systems — the responses to their actions take a form based on the processes at work, with the back-and-forth starting to reveal some of the contours. This would be good for a system, like Minstrel, that has processes of an intriguing shape (whereas it’s bad for systems like Eliza, which have basically boring underlying processes).

Right now, the best way I can think of to create an interactive media experience around something like Minstrel is to offload the commonsense reasoning on the audience. Rather than presenting finished stories generated by the Minstrel-style system, the system would be presented as a tool or partner, generating candidate fictional elements which the audience could selectively accept or reject, allowing the audience to guide the results toward the traditional or the surreal. I’m finding myself intrigued by the idea…

We can leave out unrelated knowledge (e.g., sewing doesn’t matter to the King Arthur domain) and lessen the chances of inappropriate generations.

Imagine that our cognitive processes have evolved to be very efficient for normal day-to-day situations. The picture we might have of “normal” cognition would be robust “common sense” processes operating on very relevant knowledge. So when we’re out for a walk and see someone hurl a stone at us, we can react quickly, using simple, fast cognition about how moving objects generally behave, the effects of getting hit by objects and so on. We’re not thinking about sewing, or elephants or anything else not obviously relevant.

Now we might also evolve different cognitive processes for dealing with unusual situations, particularly where our normal processes have failed. We’d expect these to be less efficient (in the sense of more often creating poor solutions) and to make use of less relevant knowledge (since we only fall back to these processes when using the relevant knowledge has failed).

If we call that second category “creativity” then it seems likely that building a computer model of creativity that uses only a very limited set of relevant knowledge in a microworld is self-defeating. And I think that’s a legitimate problem for Minstrel. To some extent you can get some interesting cross-fertilization even with a limited microworld, but it would be very interesting to see if a model could be built that could robustly reason creatively across very disparate domains.

However, the greater the leaps the system can produce (a knight eating a princess, suicide by sewing) the more we need to be able to reason about the appropriateness of what it produces.

If your criteria are something else, how to write them down is a whole different problem, though I’ll agree that if they turn out to be impossible to write down for a particular domain it’d point away from Minstrel being very useful in that domain.

Ah, but as Dave Jefferson challenged me on this issue: If you have the knowledge to recognize a bad solution, why did you create it in the first place? It’s hard to swallow a model of cognition along the lines of “generate a bunch of random answers and then filter out the ones that don’t work.” My answer goes back to the notion above about common-sense reasoning versus creativity reasoning, but I think it is a legitimate issue. And even if people do this sort of post-hoc editing of creative reasoning (and IIRC the psychologists are split on the issue), why should we build computer programs with the same limitations?

As I recall, one reason I had the post-hoc assessment in the case of device invention was that it let me re-use the TRAMs from storytelling essentially unchanged. If we borrow a notion from CBR that knowledge starts out very specific and becomes generalized as we experience it more often, then it may be that our creative processes are fairly specific (since we use creativity much less frequently than common-sense reasoning). But (as pointed out above) this is inconsistent with the notion that creativity must apply broadly across domains! So post-hoc assessment by common-sense reasoning might be a “hack” to compensate for this problem. This notion suggest some interesting experiments about people who are creative in several different domains. Have they generalized their creativity? Or are they especially good at applying their common-sense filters?

…it’s difficult to find a way for a machine to distinguish “surprising but interpretable (or even clever)” from “completely nonsensical and seemingly random”

In the case of Minstrel, it didn’t generate nonsensical ideas because everything was goal-driven. Killing yourself with a sewing needle might be naive, but it isn’t nonsensical. I always thought that Lenat’s EURISKO was interesting in this way — it couldn’t create a nonsensical solution, but it could creatively explore the solution space to find unlikely but effective solutions.

(I haven’t gotten myself a copy of the dissertation yet, unfortunately, so everything I say is from reading the book.)

The Tech Report Noah references above is the dissertation (in a more compact form).

In the more general narrative case I’ll agree it’s harder, since it’s basically a commonsense-reasoning problem to be able to reject stories that seem random, aren’t interesting, aren’t interpretable, etc.. Michael and I have been running up against a similar problem in our game-generation work: even in simple games of the level of complexity of a WarioWare microgame, it’s difficult to find a way for a machine to distinguish “surprising but interpretable (or even clever)” from “completely nonsensical and seemingly random”. Even the largest commonsense-reasoning databases (like Cyc) are still far off from encoding enough relevant cultural and cognitive knowledge, imo, to be able to distinguish that sort of thing. I’m also skeptical that some general non-messy solution exists, since I think a lot of the judgments are based on bits of cultural history and human perception that didn’t have to be the case, rather than on universal truths about rationality or something. For example, a common interpretability strategy is to reference well-known game or film or story or comic tropes, but doing so effectively requires a system to know what such tropes are, and how to know when a person would recognize them (versus the reference being too subtle, or too obscure, say)—basically a huge knowledge-engineering problem in a poorly-mapped-out space.

From the Minstrel perspective though I think that might be focusing on a criticism of something that wasn’t really the intent. I read Minstrel mainly as a proposal for how a creative system should generate things given an appropriate assessment—the whole approach of failure-driven reasoning and whatnot depends on having some notion of what a failure is. To that end I saw the “boredom” heuristic more as a really simple test case than as the main result: it demonstrates shows that if your main criterion is avoiding boredom in this toy story world, then Minstrel finds a way to do so. If your criteria are something else, how to write them down is a whole different problem, though I’ll agree that if they turn out to be impossible to write down for a particular domain it’d point away from Minstrel being very useful in that domain.

(I haven’t gotten myself a copy of the dissertation yet, unfortunately, so everything I say is from reading the book.)

In parallel I’ll go back and take a look at the passage you’re discussing. Recently I’ve also been looking at Scott’s dissertation, which goes into more detail (in some areas) than the book — so I’ll also check out the version there.

But, of course, the system building will be guided by an idea of what might work. Someone needs to have an “aha” moment — an insight into how you could keep a Minstrel-like approach freewheeling enough to create interesting results while somehow limiting inappropriate results. It might not seem as creative, in this new version, but it could serve the purpose of media-making (rather than exploring questions of computational creativity).

One possibility for this media-focused system might be to constrain the world in further respects. Another might be to somehow limit the freedom of movement given to TRAMs. A third might be some evaluation at the end that disposes of inappropriate generations.

At the time I wrote my comment I was thinking of only the last of these approaches. But perhaps the last is also the least tractable. I’m guessing it’s more likely that the first, second, or a combination of them might provide a realizable “aha” moment for someone out there.

Noah, I wonder if even within a limited domain, e.g. a game microworld, that a system like Minstrel can create some creative, unexpected, and entertaining results. If one adds a layer of reasoning about appropriateness to a Minstrel-like approach, as you suggest, I have optimism it can be used to give narrative intelligence to games / interactive stories. It would require significant R&D to achieve this, of course.

effort shifted from storytelling to creativity. Storytelling went from being an end in itself to being the domain in which Minstrel demonstrated creativity.

It’s also helpful for me that you frame questions about Minstrel in terms of the common-sense reasoning problem. Yes, limiting the domain to a microworld is a common way to try to get around this problem. In some ways this points to an exciting future for these techniques in areas of media (e.g., games). After all, games are microworlds — we can specify exactly how they behave, what objects exist, and so on. We can leave out unrelated knowledge (e.g., sewing doesn’t matter to the King Arthur domain) and lessen the chances of inappropriate generations.

However, the greater the leaps the system can produce (a knight eating a princess, suicide by sewing) the more we need to be able to reason about the appropriateness of what it produces. Which brings us to a reasoning problem that we can’t solve by limiting the domain — because the very nature of these leaps is that they stretch beyond the bounds of the expected within the domain. This means that Minstrel probably doesn’t point in the right direction for those seeking new routes to making media, even if it represents an exciting result in the area of computational creativity.

Or, at least, that’s how my thinking runs at the moment. I’d be very interested to hear if your thinking runs similarly.

Less generally, I’m also interested to read that, “At the time I began my work on Minstrel the only previous work on storytelling was Meehan’s Talespin.” Would that mean you don’t really count Natalie Dehn’s work? Reading back over the conversation from last year, I was reminded by Michael Joyce’s comment that her ideas influenced a number of people in the early 1980s.