CONSISTENCY AT THE CORE OF COMMONSENSE

Donald Perlis

Computer Science Department, University of Maryland, College Park MD, Maryland, U.S.A.

Keywords: Knowledge base, Inconsistency, Autonomous error-recovery, Reasoning, Commonsense.

Abstract: This paper argues for a "commonsense core" hypothesis, with emphasis on the issue of consistency in agent

knowledge bases. This is part of a long-term research program, in which the hypothesis itself is being

gradually refined, in light of various sorts of evidence. The gist is that a commonsense reasoning agent that

would otherwise become incapacitated in the presence of inconsistent data may – by means of a modest

additional error-handling “core” component – carry out more effective real-time reasoning, and also that

there may be cases of interest in which the “core” is more usefully integrated into the knowledge base itself.

1 INTRODUCTION

The idea of a knowledge (or belief, or information)

base (KB) is central to artificial intelligence.

Somehow, an automated agent is to make – and

possibly act upon – inferences (such as answers to

queries, or plans to achieve goals); and this inferring

makes use of whatever information (i.e., whatever

KB) may be at the agent's disposal. A great deal of

work has gone into characterizing such inferences;

most of it assumes that the KB itself is consistent.

(In what follows, I shall take “KB” to refer to a

dynamic store of beliefs together with a knowledge-

representation and reasoning framework.)

The consistency assumption has various

advantages, both theoretical and practical. There are

many important uses of logics for which the axioms

form a consistent set; indeed, this is the standard

situation in the history of logic formalisms,

including many new logics invented for use in

computer science (non-monotonic logics, temporal

logics, and so on).

But increasingly there has been interest in the

alternative so-called paraconsistent situation: the set

of axioms is inconsistent and yet – by some key

variation on classical logic – the (paraconsistent)

logic behaves usefully. We will not attempt to

review this extensive literature here, except to point

out that one motivation for this interest comes from

artificial intelligence, where the need to reason

within inconsistent knowledge bases is a serious

reality in many situations. Indeed, there is reason to

think that inconsistency is nearly inevitable in large-

scale dynamic real-world knowledge bases, and

especially in the domain of commonsense agent

behavior; see (Anderson, Gomaa, et al, 2008;

Anderson, Fults, et al 2008; Grant, 1978; Land &

Marquis, 2010; Perlis, 1997) for a sampling of work

in this area. Essentially the idea is that mistakes will

crop up, leading to eventual contradictions, and this

in turn requires a repair mechanism to avoid havoc.

The most notorious form of such havoc is this:

reasoning (inference) in classical logics with an

inconsistent set of axioms is easy; all too easy. Any

formula whatsoever is a theorem, in virtue of the

validity of the “ex contradictione quodlibet”

inference schema: from A and -A, infer B. This

property of classical (and some other) logics goes by

the aptly colorful name of “explosivity” (Priest,

1996).

The rest of this paper is organized as follows: we

describe a notion of commonsense agent behavior

for which consistency is an implausible luxury; then

our “core” hypothesis is described, along with

various forms of evidence; we conclude with a

discussion of whether the core is best viewed as part

of the KB, or as a separate module.

2 COMMONSENSE BEHAVIOR

Humans survive (by and large) in a complex and

rapidly changing world. Much of our competence is

surely due to many finely honed “instincts” suited to

distinct specialized circumstances; but we also tend

to do well when faced with highly novel or irregular

578

Perlis D..

CONSISTENCY AT THE CORE OF COMMONSENSE.

DOI: 10.5220/0003158405780581

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 578-581

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

situations requiring action and yet for which we

have no specific (instinctive or learned) responses.

Our ability not to fall apart or – stated more

positively – to carry on in an effective manner even

when things are not quite as we are used to, perhaps

captures much of what is loosely termed

commonsense; here we shall call this ability

“commonsense behavior”. To put it yet another way:

when things are suddenly amiss and a (possibly

quick) irregularity-fix is needed (but not already at

hand), we often come up with something that allows

us to continue making useful progress toward at

least some goals and to avoid huge increase in costs

(often, but not always: the financial wizards did not

manage this as the derivatives market began to

collapse, nor the oil rig engineers in the Gulf of

Mexico, nor the reactor engineers at Chernobyl).

There is an enormous AI literature on

commonsense reasoning; here we have defined

commonsense behavior to be a little broader, in that

it need not involve reasoning, or at least not subtle

reasoning used to solve tricky puzzles. Here is an

example: you are playing in an outdoor checkers

tournament, but several of your pieces fall and roll

into a storm sewer. You reach, but fail to retrieve

them. You could now ponder at length, treating this

as a logic puzzle, hoping for a special insight as to

how to retrieve the pieces. Or, you could realize that

this might take a long time, that in this situation

what matters is not those fallen pieces but rather the

ongoing tournament, and that you can ask the

referees for advice.

Now, this – asking for advice – is itself a ready-

to-hand technique, and so in a sense we do have a

fix for many novel things, as long as an expert is

handy. But using this kind of fix is very different

from knowing specifics about a particular situation.

It involves awareness of ones failing efforts, of ones

lack of a ready solution or even a good chance at

personally finding an appropriate one, and of the

availability of someone who might help. These have

less to do with checkers than with general ways of

coping with irregularities. Thus, roughly speaking,

we might divide data in a KB into that portion

relevant to a highly specific task-at-hand, and

general irregularity-fix data that might be applied

more or less independent of the situation.

Just to clarify that asking for help is not the only

such strategy, here are a few other frequently-handy

irregularity-fix strategies: try again; make small

random changes; give up. Yes: even giving up is

often a very good thing to do; certainly much better

than struggling on and on indefinitely if the cost is

great and there is little indication that success is

likely. I’ll mention one more strategy (often highly

effective, but not quick): initiate a training regimen

in order to improve (or learn) a particular behavior

whose lack was impeding progress.

Now, what has all this to do with inconsistency

in a KB? It is this: knowing that the situation is

irregular – that one does not already have a strategy

at hand – amounts to noting a mismatch between the

actual situation and anything one expects. The

agent’s KB has the item B, perhaps since the agent

has the expectation (belief) that the situation is one

where B holds; and also in the KB is –B, perhaps as

observed data. The two are in contradiction, and the

agent must treat this not as a run-of-the-mill set of

beliefs with which to reason (that would be to

blindly brook explosivity), but as a case where the

KB itself is to be looked at as a puzzle: what to do

about the anomaly of both B and –B being there?

Seen in those terms, a new task arises at a meta-level

(that of KB management) and the contradiction

becomes a possibly important clue to something

needing attention rather than a logical nuisance.

3 CORE HYPOTHESIS

We make the following hypothesis: there is a set of

general-purpose irregularity-fix strategies that is:

 adequate to a broad range of novel situations

 concise

 implementable

 largely independent of the size of the KB

overall or complexity of the task or domain

 consistent in itself

We refer to this as the commonsense core

hypothesis. It can be considered as postulating a

fragment of an agent’s inferential capacity that has

the above properties. Here we will very briefly

mention some evidence in its favor.

First of all, humans seem firmly possessed of

just such a core set of irregularity-fix strategies. We

are quite marvellous at dealing on the fly with

myriad unanticipated situations, relying on the same

few general-purpose techniques over and over (such

as those listed earlier). And – in case that intuitive

claim is not convincing – controlled empirical

studies have produced data on such strategies in

laboratory settings where subjects make high-level

general judgments as to their progress vis-à-vis time

remaining, confidence in their work, expressions

they do not understand, and so on; see (Nelson &

Dunlosky, 1994; Nelson, Dunlosky, et al, 1994).

Further, neurocognitive work suggests particular

CONSISTENCY AT THE CORE OF COMMONSENSE

579

brain structures implicated in error-noting; see

(Kendler & Kendler, 1962, 1969). Finally, various

implementations of the above ideas have shown

success in a wide range of domains (Anderson,

Fults, et al, 2008); In these cases however the “core”

of irregularity-fix strategies was separate from the

system’s KB.

Irregularity-fix cores for autonomous agents

does exist, as the afore-mentioned implementations

show. But can they – or some improved future

version – do all that has been hypothesized? What

sorts of irregularities might lie beyond any given

core set of fixes? Are we in a Godelian situation

where, for any core set, there are yet more

irregularities beyond its reach? And if a core can

reach far, will it no longer be concise? Will core

effectiveness scale with the KB? Will a powerful

core also require a powerfully expressive language

and possibly thereby risk inconsistencies within

itself?

There are grounds to think that reach and

conciseness and effectiveness are well within the

capacity of an implementable commonsense core:

much the same grounds cited above for the existence

of a core in the first place. But scalability? As

humans are faced with more and more information,

our effectiveness can sometimes degrade in two

ways. Not only can it take us longer consider all the

data (though in some cases of course, the extra

information makes things go faster), but also there is

a heightened likelihood that we will mess up: we’ll

forget something, lose track of where we are,

confuse or conflate similar notions, etc. However,

this is not the issue; rather it is whether we cope as

well, whether we still notice things amiss and bring

corrective strategies to bear, as well as we do when

we have a smaller set of facts to deal with. Here I

simply state an opinion (in the current absence of

empirical data): yes, we do notice our confusion, our

lack of progress, and so on, whether working with a

large or small KB, on a simple or complex task, and

we also respond actively as well: we start over, or

ask for help, give up, etc. But we do not rotely go on

and on oblivious to the mess we are in.

4 SHOULD THE CORE FIT

INTO THE KB?

We now address the last hypothesized item:

consistency within the core. As claimed, the

commonsense core can be implemented and

included as part of an autonomous system. Having

the core sit outside the KB – for instance as a

monitor-and-control Bayesian net apart from the

agent’s world model – is an effective design for

many purposes. Further, its isolation then protects it

from possible infection from a contradiction in the

KB. While the KB may be in the throes of explosive

inference, the core is not. Even the beginnings of an

explosive KB inference process are readily noted by

such a core, which in turn then can redirect KB

inference in more productive ways. If the core fixes

are expressed in propositional language, and

together form a concise set, and if each fix is of the

simple sort we have described (ask for help, give up,

etc) it is plausible that there may well be no internal

inconsistency between them.

Yet there are situations in which it may make

less sense to separate the core from the KB. Here are

four such situations: (i) over time the core trains a

new item into the KB so that what had been a

particular kind of anomaly handled directly by the

core becomes encoded as a familiar event: the core

strategy that had been handling these events is now

largely replicated in the KB as a standard piece of

knowledge about how the world works; (ii) the

query “why did you do that?” may require reference

to the core, and so the KB reasoner must have some

ability to monitor facts about the core: “I did that

because I got confused and had to start over”; (iii)

“how/why did I do that?” can be asked as an

exercise in self-improvement (maybe it can be done

better), which suggests bidirectional monitoring and

control between core and KB; and (iv) the core itself

may behave in an anomalous manner (and if an

infinite regress of anomaly-handling meta-cores is to

be avoided then we might as well have the all the

anomaly-handling inside a single KB at the outset).

On the other hand, combining core and KB raises

the danger of inconsistency infecting the core; how

serious a problem this may be is currently under

investigation.

REFERENCES

Anderson, M., Gomaa, W., Grant, J., Perlis, D., 2008a.

Active logic semantics for a single agent in a static

world. Artificial Intelligence 172: 1045-63.

Anderson, M., Fults, S., Josyula, D., Oates, T., Perlis, D.,

Schmill, M., Wilson, S., Wright, D., 2008b. A self-

help guide for autonomous systems. AI Magazine,

29(2):67-76.

Grant, J., 1978 Classifications for inconsistent theories.

Notre Dame Journal of Formal Logic, 3: 435-444.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

580

Lang, J., Marquis, P., 2010. Reasoning under

inconsistency: A forgetting-based approach. Artificial

Intelligence, 174: 799-823.

Nelson, T, Dunlosky, J., 1994. Norms of paired-associate

recall during multitrial learning of Swahili-Engush

translation equivalents. Memory, 2:325-335.

Nelson, T., Dunlosky, J., Graf, A., Narens, L., 1994.

Utilization of metacognitive judgments in the

allocation of study during multitrial learning.

Psychological Science, 4: 207-213.

Kendler, H., Kendler, T., 1962. Vertical and horizontal

processes in problem solving. Psychological Review,

69: 16

Kendler, H., Kendler, T., 1969. Reversal-shift behavior:

some basic issues. Psychological Bulletin, 72:229-

232.

Perlis, D., 1997. Sources of, and exploiting, inconsistency:

preliminary report. Jounral of Applied Non-classical

Logics, 7:13-24.

Priest, G., 1996. Paraconsistent logic. In Stanford

Encyclopedia of Philosophy (online version: http://

plato.stanford.edu/entries/logic-paraconsistent/ )

CONSISTENCY AT THE CORE OF COMMONSENSE

581