Explaining Explaining

Sergei Nirenburg

, Marjorie McShane

, Kenneth W. Goodman

and Sanjay Oruganti

Department of Cognitive Science, Rensselaer Polytechnic Institute, 110 8

St. Troy, NY 12180-3590, U.S.A.

Institute for Bioethics and Health Policy, Miller School of Medicine, University of Miami,

P.O. Box 016960 (M-825) Miami, FL 33101, U.S.A.

Keywords: Explanation, Explainable AI, Cognitive Modeling.

Abstract: Explanation is key to people having confidence in high-stakes AI systems. However, machine-learning-based

systems – which account for almost all current AI – can’t explain because they are usually black boxes. The

explainable AI (XAI) movement hedges this problem by redefining “explanation”. The human-centered

explainable AI (HCXAI) movement identifies the explanation-oriented needs of users but can’t fulfill them

because of its commitment to machine learning. In order to achieve the kinds of explanations needed by real

people operating in critical domains, we must rethink how to approach AI. We describe a hybrid approach to

developing cognitive agents that uses a knowledge-based infrastructure supplemented by data obtained

through machine learning when applicable. These agents will serve as assistants to humans who will bear

ultimate responsibility for the decisions and actions of the human-robot team. We illustrate the explanatory

potential of such agents using the under-the-hood panels of a demonstration system in which a team of

simulated robots collaborates on search task assigned by a human.

1 INTRODUCTION

Explanation is clearly one of Marvin Minsky’s

“suitcase” words “that we use to conceal the

complexity of very large ranges of different things

whose relationships we don’t yet comprehend”

(Minsky, 2006, p. 17). The Stanford Encyclopedia of

Philosophy (https://plato.stanford.edu/) includes

detailed entries on mathematical, metaphysical, and

scientific explanation, and a separate one on causal

approaches to the latter. Specialist philosophical

literature discusses Carl Hempel’s (1965) deductive-

nomological model of explanation and the rival

inductive-statistical approaches. Explanation is also

discussed in other disciplines, such as psychology

(e.g., Lombrozo, 2010). Special attention is also paid

to the differences between explainability,

interpretability, transparency, explicitness, and

faithfulness (e.g., Rosenfeld & Richardson, 2019).

Recent years have also seen a pronounced interest in

developing novel theories of explanation (Yang,

Folke, & Shafto, 2022; Rizzo et al., 2023).

Explanation in AI has a long history as well.

Arguably the first AI-related contribution was Craik

(1943). Kenneth Craik was a psychologist and an

early cyberneticist whose work influenced AI and

cognitive science (Boden, 2006, pp. 210-218). His

book, entitled “The Nature of Explanation,” discusses

a variety of the senses of this suitcase word and,

among other things, stresses the distinction between

causal explanation in terms of a formal world model

(what would later be termed ontology) and statistical

explanation, which seeks to explain by pointing out

uninterpreted relations among observable entities.

The above distinction still remains in the spotlight

today. Most current generative AI systems are black

boxes whose functioning cannot be explained in

normal human terms. For certain applications, this is

not a problem:

1. Non-critical AI capabilities – such as GPS

systems, machine translation systems, and

search engines – are widely and happily

employed by end users who don’t require

explanations.

2. AI capabilities that emulate physical rather than

cognitive capabilities – such as robotic movement

and speech recognition – are incompatible with

the kinds of user-elucidating explanations we

address here. That is, we all understand that it

would be pointless to ask a robot exactly how it

extends its arm or keeps its balance when walking

on uneven surfaces.

116

Nirenburg, S., McShane, M., Goodman, K. and Oruganti, S.

Explaining Explaining.

DOI: 10.5220/0013000600003886

In Proceedings of the 1st International Conference on Explainable AI for Neural and Symbolic Methods (EXPLAINS 2024), pages 116-123

ISBN: 978-989-758-720-7

3. AI systems that emulate cognitive

capabilities (e.g., X-ray analysis systems)

can be useful, despite their lack of

explainability, as long as they are leveraged

as orthotic assistants rather than autonomous

systems. This requires that domain experts

understand the reliability and deficits of the

technology well enough to judiciously

incorporate it into their workflow (Chan &

Siegel, 2019; Nirenburg, 2017).

By contrast, lack of explainability is a problem for

ML-based AI systems that are designed to operate

autonomously in critical domains. For example, as of

June 2021, the FDA cleared 343 AI/ML-based

medical devices, with over 80% of clearances

occurring after 2018 (Matzkin, 2021). This supply of

new AI systems has continued unabated even though

their adoption has been less than enthusiastic. Fully

70% of these devices offer radiological diagnostics

and typically claim to exceed the precision and

efficiency of humans. But, according to Gary Marcus

(2022), as of March 2022, “not a single radiologist

has been replaced.” So, regulators keep approving

systems whose operation cannot be explained, and

developers keep hoping that their systems, though

unexplainable, will be adopted. (For further

discussion, see McShane, Nirenburg, & English,

2024, Section 2.7.1.)

The unexplainability problem has been addressed

in earnest by the Explainable-AI (XAI) movement

(Finzel et al., 2023; Bodria et al., 2021; Cambria et al.,

2022; Nagahisarchoghaei et al., 2023; Schwalbe &

Finzel, 2024), but the results are, in general, less than

satisfying (Barredo et al., 2020). XAI investigators

hedge the explainability problem by redefining

“explanation” (Gunning, 2017; Mueller et al., 2019).

XAI research does not seek to explain how systems

arrived at their output. Instead, it concentrates on

“post hoc algorithmically generated rationales of

black-box predictions, which are not necessarily the

actual reasons behind those predictions or related

causally to them... [and which] are unlikely to

contribute to our understanding of [a system’s] inner

workings” (Babic et al., 2021).

The related human-centered explainable AI

(HCXAI) movement, for its part, identifies the

explanation-oriented needs of users but is hampered

in fulfilling them because of its commitment to

machine learning (Babic et al., 2021; Ehsan et al.,

2022; Liao & Varshney, 2022).

The solution to the problem of unexplainable ML-

based AI is not to keep trying to square that circle:

generative AI techniques are, and will remain,

unexplainable. The solution is to step back and

reconsider how to develop AI systems so that they are

explainable to the degree, and in the ways, that are

necessary for different types of applications.

Figure 1: The demonstration view of the robotic search-and-retrieve system at the end of the system run. The simulation

window is to the left, the dialog window is in the middle, and four of many available under-the-hood panels are to the right.

This figure is intended only to show the look-and-feel of the system; subsequent figures of individual components will be

legible.

Explaining Explaining

117

Intelligent behavior by humans and AI agents

involves a variety of different capabilities of

perception, reasoning, decision-making, and action.

Some of them are arguably better fit to be

implemented using generative AI approaches, some

others, by symbolic AI approaches. Therefore, hybrid

AI systems are better suited for comprehensive (non-

silo) applications than either of the above approaches

alone. This observation was first made over 60 years

ago (Minsky, 1961) and has finally received due

attention in the field. Indeed, hybrid “neurosymbolic”

AI architectures are at present one of the foci of work

in AI (Hitzler et al., 2023).

Our team is developing a family of hybrid

cognitive systems that we call LEIAs: Language-

Endowed Intelligent Agents. Our work is a part of the

movement toward integrating empirical (deep

learning-based) and deductive/rational (knowledge-

based) approaches to building intelligent agent

systems. Explanation is an important component of

such systems (de Graaf et al., 2021).

We believe that explanations are needed, first and

foremost, for those aspects of agent functioning that

end users want to be explained, so that they will come

to trust their agent collaborators. (Granted,

explanations are also useful for AI developers. But

enhancing the efficiency of system development is a

different concern from supplying systems with

essential explanatory capabilities.) It follows that

generative AI methods are most useful for

implementing system modules that don’t need to be

explained, such as motor control and uninterpreted

perception. For any process that requires explainable

reasoning (perception interpretation, decision-

making, action specification, etc.), and for any

application where confidence in system output is

important, black-box methods, such as large language

models (LLM), are not a good fit. Accordingly,

LEIAs use LLMs whenever this simplifies or speeds

up work on tasks that do not require explanation or

for which errors will not undermine the reliability of

agent behavior.

In what follows we briefly illustrate some of the

explanatory capabilities of LEIAs.

2 LEIAs

The LEIA program of R&D is a theoretically

grounded, long-term, effort that has two main

emphases: developing cognitive robotic systems

whose capabilities extend beyond what machine

learning alone can offer, and earning people’s trust in

those systems through explainability (McShane,

Nirenburg, & English, 2024).

LEIAs are implemented in a dual-control

cognitive-robotic architecture that integrates

strategic, cognitive-level decision-making with

tactical, skill-level robot control (Oruganti et al.,

2024). The strategic (cognitive) layer relies primarily

on knowledge-based computational cognitive

modeling for interpreting perceptive inputs,

reasoning, decision-making, learning, etc. The

tactical (robotic, skill-level) module relies on data-

driven tools for recognizing perceptive inputs and

rendering actions.

LEIAs can explain their operation because they

are configured using human-inspired computational

cognitive modeling. Their explanations make clear

the relative contributions of symbolic and data-driven

methods, which is similar to a human doctor

explaining a recommended procedure using both

causal chains, such as how the procedure works, and

population-level statistics, such as the percentage of

patients for whom it is curative.

3 EXPLANATION VIA

UNDER-THE-HOOD PANELS

As detailed in McShane, Nirenburg, and English

(2024, chapter 8, “Explaining”), there are many

things that a LEIA can explain (what it knows, how it

interpreted an input, why it made a given decision,

etc.) and there are many ways to present explanations

to people. Although the most obvious way is through

language, other expressive means can be even more

useful in some contexts. One such way is by

dynamically showing traces of system operation

using what we call under-the-hood panels.

We first introduced under-the-hood panels in the

Maryland Virtual Patient (MVP) proof-of-concept

clinician training application (McShane et al., 2008;

McShane & Nirenburg, 2021). There, the under-the-

hood panels showed traces of the physiological

simulation of the virtual patient, the patient’s

interoception, its thoughts, the knowledge it learned,

and how it interpreted text inputs from the user, who

was playing the role of attending physician. These

insights into system functioning were geared toward

earning the trust of medical educators, who would

ultimately need to choose to incorporate such a

system into their pedagogical toolbox.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

118

3.1 A Search-and-Retrieve Request

We will illustrate the explanatory power of under-the-

hood panels using a new system (Oruganti et al.,

2024; Oruganti et al., submitted) in which two

simulated robots, a drone and a ground vehicle, work

as a team to fulfill a search-and-retrieve request by a

person (Figure 1). A human named Danny, who is

located remotely, asks the team of robots – a drone

and ground vehicle (UGV) – to find keys that he lost

in his apartment. Danny communicates with the

UGV, since it is serving as the robotic team leader

with the drone as its subordinate. The full dialog,

which we’ll walk through, is shown in Figure 2.

Figure 2: The full dialog in the demo. The prompt for typing

is at the bottom.

When Danny issues his request, both robots

semantically interpret the input, resulting in text-

meaning representations (TMRs) that are written in

the ontologically-grounded metalanguage used for all

agent knowledge, memory, and reasoning. This

metalanguage uses unambiguous ontological

concepts (not words of English) and their instances,

described by ontologically-grounded properties and

values. Ontological concepts are written in small caps

to distinguish them from words of English, and their

instances are indicated by numerical suffixes. The

process of language understanding is complicated, as

detailed in McShane and Nirenburg (2021).

Figure 3 shows the UGV’s TMR of Danny’s

request and Figure 4 shows its subsequent thoughts,

which are natural language traces of its reasoning for

the benefit of humans. (It reasons in the ontological

metalanguage.)

Figure 3: The UGV's interpretation of Danny's request.

Figure 4: The UGV's thoughts in response to Danny's

request.

Because the UGV has received a request for

action, and because it knows that has a helper (the

drone), it places a

COLLABORATIVE

ACTIVITY

on its

Agenda (Figure 5). Before it launches the plan for

FOR

LOST

OBJECT

it has to check if its

preconditions are met (Figure 5). The first

precondition, knowing the object type, is already met

(i.e., keys), but the second and third are not: knowing

the keys’ features and knowing where they were last

seen; so the UGV asks about these things in turn. The

reasoning associated with this sequence of actions is

shown in Figure 6.

Explaining Explaining

119

Figure 5: The UGV's agenda while it is fulfilling

preconditions for searching for the keys.

Figure 6: The UGV’s thoughts as it fulfilling preconditions

for searching for the keys.

The UGV then proposes the plan of searching the

apartment to the drone, the drone agrees, and it

launches a plan to do that. Its thoughts – including

those running up to this move – are shown in Figure 7.

Figure 7: The drone's thoughts leading up to and including

its starting to search the apartment.

Since the robots in this simulation are decentralized,

each having its own cognitive layer, the drone

independently carries out much of the same reasoning

as the UGV. (Note that our architecture also permits

centralized robots that share a cognitive layer.)

Having agreed upon a plan, the UGV and the

drone leave their charging stations, highlighted in

green in Figure 8, and begin searching the apartment.

Figure 8: The room layout just before the robots being

searching, with the robots and keys shown in green boxes.

The robots are equipped with sensors to detect,

identify, and tag objects, and to combine this data to

localize objects and themselves. Interpreted traces of

what they are seeing are shown in visual meaning

representations (VMRs) in the associated under-the-

hood panel. VMRs are similar in form and content to

TMRs since, no matter what an agent perceives or

which mode of perception it uses (hearing, vision,

etc.), it has to interpret the stimuli into ontologically-

grounded knowledge that feeds its reasoning. Figure

9 shows a moment when the UGV is looking at a

particular spot on the blue-green striped carpet.

Figure 9: A visual meaning representation (VMR).

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

120

The robots engage in a search strategy involving

waypoints, zones and sub-zones that are pre-

designated for the apartment environment (Oruganti

et al., 2024). The search action is triggered through

action commands from the strategic layer but the

search itself is controlled by the tactical (robotic)

layer. The cognitive (strategic) module knows which

zones exist but does not guide how the robots

maneuver through those zones. The simulation

system is equipped with timing strategies and

modules to ensure process and data synchronization

between the tactical and strategic layers.

Searching each zone is a subtask of the plan

FIND

LOST

OBJECT

. After completing each subtask – i.e.,

searching each zone – each robot reports to the other

one about whether it was successful, which is driven

by the

COLLABORATIVE

ACTIVITY

plan.

When the team leader finds the keys, it ceases

searching and first reports this to its subordinate and

then to Danny. The trace of this reasoning is shown in

Figure 9. It uses different formulations for each of

them because its language generation system (whose

traces are not shown in this demo system) is designed

to mindread its interlocutors and present information

in the most useful way for them. Whereas these robots

operate in terms of cardinal directions, making north

of the couch a good descriptor, most humans prefer

relative spatial terms like behind the couch.

Figure 10: The UGV's thoughts when it finds the keys and

decides to report that.

4 CONCLUSIONS

Explainability is essential to critical applications, and

in order for systems to be truly explanatory, they must

first of all understand what they are doing. This

requires that they be grounded in high-quality

knowledge bases that optimally integrate causal and

correlational reasoning.

This paper focused on explanation via traces of

system operation using under-the-hood panels. The

panels selected for this demo displayed the agents’

interpretation of language inputs and visual stimuli,

their reasoning, and their agenda. Much more could

be shown if target users would find that helpful: the

agents’ ontologies, episodic memories, lexicons,

decision-making about language generation, and so

on. The under-the-hood panels do not attempt to

capture unexplainables that are implemented using

machine learning, such as what drives robotic

movement or the robots’ approach to searching a

space.

In the current benchmark-driven climate, under-

the-hood panels offer an alternative standard of

system evaluation.

Under-the-hood panels are just one mode of

explanation for LEIAs. The other primary one is

language. The many things that a LEIA can explain

using language are detailed in Chapter 8 of McShane,

Nirenburg, and English (2024).

Although the theoretical, methodological, and

knowledge prerequisites for explanation by LEIAs

are quite mature, this doesn’t mean that all problems

associated with explanation are solved.

Consider the example of physicians explaining

relevant aspects of clinical medicine to patients, a

capability that was relevant for the MVP clinician-

training system mentioned above. The task has two

parts: deciding what to say and how to say it. Both of

these depend not only medical and clinical

knowledge, but also on the salient features of patients,

such as their health literacy (as hypothesized by the

physician), their interest in medical details, their

ability to process information based on their physical,

mental, and emotional states, and so on. Identifying

these salient features involves mindreading

(Spaulding, 2020) – also known as mental model

ascription. For example, an explanation may be

presented in many different ways:

• as a causal chain: “You feel tired because of an

iron deficiency.”

• as a counterfactual argument: “If you hadn’t

stopped taking your iron supplement you

wouldn’t be feeling so tired.”

• as an analogy: “Most people find it easier to

remember to take their medicine first thing in

the morning; you should try that.”

• using a future-oriented mode of explanation:

“If you take your iron supplement regularly,

you should feel much more energetic.”

Explaining Explaining

121

Moreover, explanations are not limited to speech

– they can include images, videos, body language,

live demonstration, and more. Overall, generating

explanations tailored to particular humans is a

difficult task. However, as with all other aspects of

cognitive modeling, simplified solutions hold

promise to be useful, particularly given the well-

established fact that adding more content to an

explanation does not necessarily make it better (cf.

the discussion of decision-making heuristics in

Kahneman, 2011).

ACKNOWLEDGEMENTS

This research was supported in part by Grant

#N00014-23-1-2060 from the U.S. Office of Naval

Research. Any opinions or findings expressed in this

material are those of the authors and do not

necessarily reflect the views of the Office of Naval

Research.

REFERENCES

Babic, B., Gerke, S., Evgeniou, T., Cohen, I. G. (2021).

Beware explanations from AI in health care. Science,

373(6552), 284–286.

Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., et al.

(2020). Explainable artificial intelligence (XAI):

Concepts, taxonomies, opportunities and challenges

toward responsible AI. Information Fusion, 548: 82-

115.

Boden, M. (2006). Minds as Machines. Oxford University

Press.

Bodria, F., Giannotti, F., Guidotti, R., Naretto, F.,

Pedreschi, D., Rinzivillo, S. (2021). Benchmarking and

survey of explanation methods for black box models.

Arxiv:2102.13076v1.

Cambria, E., L. Malandri, F. Mercorio, M. Mezzanzanica,

and N. Npbani. 2023. A Survey on XAI and natural

language explanations. Information Processing and

Management, 60, 103-111.

Chan, S., Siegel, E. L. (2019). Will machine learning end

the viability of radiology as a thriving medical

specialty? British Journal of Radiology, 92(1094).

https://doi.org /10.1259/bjr.20180416

Craik, K. J. W. (1943), The Nature of Explanation.

Cambridge University Press.

De Graaf, M. M. A., A. Dragan, B. F. Malle, and T. Ziemke.

2021. Introduction to the Special Issue on Explainable

Robotic Systems. J. Hum.-Robot Interact. 10, 3, Article

22. https://doi.org/10.1145/3461597

Ehsan, U., Wintersberger, P., Liao, Q. V., Watkins, E. A.,

Manger, C., Daumé, Hal, III, Riener, A., Riedl, M. O.

(2022). Human-centered explainable AI (HCXAI):

Beyond opening the black-box of AI. CHI EA ’22:

Extended abstracts of the 2022 CHI Conference on

Human Factors in Computing Systems, pp. 1–7.

Association for Computing Machinery.

https://doi.org/10.1145/3491101.3503727

Finzel, B., Schwalbe, G. (2023). A comprehensive

taxonomy for explainable artificial intelligence: a

systematic survey of surveys on methods and concepts.

Data Mining and Knowledge Discovery.

https://doi.org/10.1007/s10618-022-00867-8

Gunning, D. (2017). Explainable artificial intelligence

(XAI). DARPA/I20. Program Update November 2017.

Hempel, C. G. (1965). Aspects of scientific explanation. In

Hempel, C. G. (1965), Aspects of Scientific Explanation

and Other Essays in the Philosophy of Science. Free

Press, pp. 331– 396.

Hitzler, P., M.K. Sarker and A. Eberhart (eds). 2023.

Compendium of Neurosymbolic Artificial Intelligence.

Frontiers in Artificial Intelligence and Applications,

vol. 369. IOS Press.

Kahneman, D. (2011). Thinking: Fast and slow. Farrar,

Strauss and Giroux.

Liao, V. Q., Varshney, K. R. (2022). Human-centered

explainable AI (XAI): From algorithms to user

experiences. arXiv:2110.10790.

Lombrozo, T. (2010). Causal–explanatory pluralism: How

intentions, functions, and mechanisms influence causal

ascriptions. Cognitive Psychology, 61(4): 303–332.

doi:10.1016/j.cogpsych.2010.05.002

Marcus, G. (2022, March 10). Deep learning is hitting a

wall. Nautilus. https://nautil .us/deep-learning-is-

hitting-a-wall-14467

Matzkin, A. (2021, September 29). AI in Healthcare:

Insights from two decades of FDA approvals. Health

Advances blog.

McShane, M., Jarrell, B., Fantry, G., Nirenburg, S., Beale,

S., Johnson, B. (2008). Revealing the conceptual

substrate of biomedical cognitive models to the wider

community. In J. D. Westwood, R. S. Haluck, H. M.

Hoffman, G. T. Mogel, R. Phillips, R. A. Robb, & K.

G. Vosburgh (Eds.), Medicine meets virtual reality 16:

Parallel, combinatorial, convergent: NextMed by

design (pp. 281–286). IOS Press.

McShane, M., Nirenburg. S. (2021). Linguistics for the age

of AI. MIT Press. Available, open access, at

https://direct.mit.edu/books/book/5042/Linguistics-

for-the-Age-of-AI.

McShane, M., Nirenburg, S., English, J. (2024). Agents in

the Long Game of AI: Computational cognitive

modeling for trustworthy, hybrid AI. MIT Press.

Minsky, M. 1961. Steps Toward Artificial Intelligence.

Proceedings of the Institute of Radio Engineers, 49: 8–

30; reprinted in Feigenbaum and Feldman, (eds.)

Computers and Thought. McGraw-Hill. 1963.

Minsky. M. (2006). The Emotion Machine. Simon and

Schuster.

Mueller, S. T., Hoffman, R. R., Clancey, W., Emrey, A., &

Klein, G. (2019). Explanation in human-AI systems: A

literature meta-review. Synopsis of key ideas and pub-

lications, and bibliography for explainable AI.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

122

Technical Report, DARPA Explainable AI Program.

arXiv:1902.01876.

Nagahisarchoghaei, M., Nur, N., Cummins,, L., Karimi, M.

M., Nandanwar, S., Bhattacharyya, S., Rahimi, S.

(2023). An empirical survey on explainable AI

technologies: Recent trends, use-cases, and categories

from technical and application perspectives.

Electronics, 12, 1092. https://doi.org/10.3390/

electronics12051092

Nirenburg, S. (2017). Cognitive systems: Toward human-

level functionality. Special Issue on Cognitive Systems.

Artificial Intelligence Magazine, 38(4), 5–12.

https://doi .org/10.1609/aimag.v38i4.2760

Sanjay Oruganti, Sergei Nirenburg, Marjorie McShane,

Jesse English, Michael Roberts, and Christian Arndt.

“HARMONIC: A Framework for Explanatory

Cognitive Robots”. In: Proceedings of the IEEE

Conference on Robotics and Automation (ICRA@40),

Rotterdam, Netherlands. 2024

Oruganti, S.S.V., S. Nirenburg, M. McShane, J. English, M.

Roberts & C. Arndt. (Submitted). HARMONIC:

Cognitive and Control Collaboration in Human-

Robotic Teams. Submitted to ICRA-2025.

Rizzo, M, Albarelli, A., Lucchese, C., Conati, C. (2023). A

theoretical framework for AI models explainability.

Proceedings of IJCAI-23.

Rosenfeld, A., Richardson, A. (2019). Explainability in

human–agent systems. Autonomous Agents and Multi-

Agent Systems, 33: 673–705.

Schwalbe, G., Finzel, B. (2024). A comprehensive

taxonomy for explainable artificial intelligence: a

systematic survey of surveys on methods and concepts.

Data Mining and Knowledge Discovery, 38: 3043–

3101. https://doi.org/10.1007/s10618-022-00867-8

Spaulding, S. (2020). What is mindreading?

Interdisciplinary Review of Cognitive Science.

https://wires.onlinelibrary.wiley.com/doi/10.1002/wcs.

1523

Yang, S. C-H., Folke, T., & Shafto, P. (2022). A

psychological theory of explainability.

arXiv:2205.08452.

Explaining Explaining

123