=Paper=
{{Paper
|id=Vol-3208/paper5
|storemode=property
|title=Online Incremental Learning with Abstract Argumentation Frameworks
|pdfUrl=https://ceur-ws.org/Vol-3208/paper5.pdf
|volume=Vol-3208
|authors=Hamed Ayoobi,Ming Cao,Rineke Verbrugge,Bart Verheij
|dblpUrl=https://dblp.org/rec/conf/comma/Ayoobi0VV22
}}
==Online Incremental Learning with Abstract Argumentation Frameworks==
https://ceur-ws.org/Vol-3208/paper5.pdf
Online Incremental Learning with Abstract
Argumentation Frameworks⋆
H. Ayoobi1,3,∗ , M. Cao2 , R. Verbrugge1 and B. Verheij1
1
Department of Artificial Intelligence, Bernoulli Institute, Faculty of Science and Engineering, University of Groningen,
The Netherlands.
2
Institute of Engineering and Technology (ENTEG), Faculty of Science and Engineering, University of Groningen, The
Netherlands.
3
Department of Computing, Faculty of Engineering, Imperial College London, United Kingdom
Abstract
The environment around general-purpose service robots has a dynamic nature. Accordingly, even the
robot’s programmer cannot predict all the possible external failures which the robot may confront. This
research proposes an online incremental learning method that can be further used to autonomously
handle external failures originating from a change in the environment. Existing research typically
offers special-purpose solutions. Furthermore, the current incremental online learning algorithms can
not generalize well with just a few observations. In contrast, our method extracts a set of hypotheses,
which can then be used for finding the best recovery behavior at each failure state. The proposed
argumentation-based online incremental learning approach uses an abstract and bipolar argumentation
framework to extract the most relevant hypotheses and model the defeasibility relation between them.
This leads to a novel online incremental learning approach that overcomes the addressed problems
and can be used in different domains including robotic applications. We have compared our proposed
approach with state-of-the-art online incremental learning approaches and an approximation-based
reinforcement learning method. The experimental results show that our approach learns more quickly
with a lower number of observations and also has higher final precision than the other methods.
Keywords
Argumentation-Based Learning, Online Incremental Learning, Argumentation Theory, General Purpose
Service Robots
1. Introduction
This paper is a short version of our journal paper [1]. The development and application of
domestic service robots are growing rapidly. Whereas basic household robots are already
ArgML2022: 1st International Workshop on Argumentation Machine Learning @COMMA 2022, September 13, 2022,
Cardiff, Wales, UK
⋆
This work is conducted at DSSC and sponsored by a Marie Skłodowska-Curie COFUND grant, agreement no.
754315.
∗
Corresponding author.
Envelope-Open h.ayoobi@imperial.ac.uk (H. Ayoobi); m.cao@rug.nl (M. Cao); l.c.verbrugge@rug.nl (R. Verbrugge);
bart.verheij@rug.nl (B. Verheij)
GLOBE https://www.rug.nl/staff/h.ayoobi/research (H. Ayoobi); https://www.rug.nl/staff/m.cao/ (M. Cao);
https://rinekeverbrugge.nl/ (R. Verbrugge); https://www.ai.rug.nl/~verheij/ (B. Verheij)
Orcid 0000-0002-5418-6352 (H. Ayoobi); 0000-0001-5472-562X (M. Cao)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
common practice [2], the study of General Purpose Domestic Service Robots (GPSR) able to do
complex tasks is increasing [3, 4]. Due to the dynamic environment around GPSRs, they need
to efficiently handle noise and uncertainty [5].
On the hardware level of GPSRs, any kind of system failure should be avoided. On a practical
level, which involves persistent changes in the environment, it becomes much more difficult to
account for all possible external failures at design time. Therefore, it is important to note that
confronting unforeseen failures is mostly the default state for GPSRs, rather than an exceptional
state as often described in the literature. There are some solutions for external failure recovery
in the literature, which involve using simulations for the prediction of external failures [6]
and logic-based reasoning to account for external failures [7, 8]. However, in most of these
cases, the solutions are proposed for specific applications. In the following, we use the word
“Failure” instead of the word “External Failure” for conciseness. This means that the focus of our
research is not on system/hardware failures. In this paper, we propose an argumentation-based
incremental online learning method for recovering from unforeseen failures.
1.1. Argumentation
Argumentation is a reasoning model based on interaction between arguments [9]. Argumenta-
tion has been used in various applications such as non-monotonic reasoning [10], inconsistency
handling in knowledge bases [11], and decision making [12]. In [13], Dung has defined an
Abstract Argumentation Framework (AF) as a pair of the arguments (whose inner structures
are unknown) and a binary relation representing the attack relation among the arguments.
Extending Dung’s idea, some arguments can support a conclusion and others might be against
(attacking) that conclusion in the bipolar argumentation framework [14]. Both the Bipolar
Argumentation Framework (BAF) and the Abstract Argumentation Framework (AF) are used in
the proposed argumentation-based learning approach.
1.2. Argumentation in Machine Learning
According to a recent survey by Cocarascu et al. [15], the works using argumentation in
supervised learning are listed as follows. Argumentation-Based Machine Learning (ABML) [16]
uses the CN2 classification approach [17]. This method uses experts’ arguments to improve the
classification results. The paper by Amgoud et al. [18] explicitly uses argumentation. There are
other approaches for improving classification using argumentation in the literature [19].
Machine learning techniques have also been used for argumentation mining [20, 21, 22].
Bishop et al. combined argumentation with machine learning to prevent failure in deep neural
network based break-the-glass access control systems [23].
In contrast with the aforementioned methods, we do not use argumentation for improving
the current machine learning approaches or resolving conflicting decisions between current
classification methods; instead, we focus on the development of an online incremental learning
method. Moreover, the proposed method only uses class labels for the testing phase and not for
the training. Therefore, it can be utilized in open-ended (class-incremental) scenarios as well
[24].
2. Background
The Abstract Argumentation Framework (AF) and Bipolar Argumentation Framework (BAF)
are the building blocks of the online incremental learning approach proposed in this paper. AF,
BAF and online incremental machine learning algorithms are formally defined in this section.
2.1. Formal Definition of Abstract Argumentation Framework
An argumentation framework defined by Dung [13] is a pair AF = (AR, 𝑅𝑎𝑡𝑡 ) where AR is a set
of arguments, and 𝑅𝑎𝑡𝑡 is a binary relation on AR, i.e. 𝑅𝑎𝑡𝑡 ⊆ AR × AR. The meaning of A 𝑅𝑎𝑡𝑡 B is
that A attacks B where A and B are two arguments. In order to define the grounded extension
semantics in AF, which is used in the proposed learning method, some semantics should be
defined first.
(Conflict-Free) Let 𝑆 ⊆ 𝐴𝑅. S is conflict-free iff there is no 𝐵, 𝐶 ∈ 𝑆 such that B attacks C.
(Acceptability) An argument A ∈ AR is acceptable with respect to a set S of arguments iff for
each argument B ∈ AR: if B attacks A then B is attacked by at least one element of S.
(Admissibility) A conflict-free set of arguments S is admissible iff each argument in S is
acceptable with respect to S.
(Characteristic Function) The characteristic function 𝐹𝐴𝐹 in an argumentation framework AF
= (AR, 𝑅𝑎𝑡𝑡 ) is defined as follows:
𝐹𝐴𝐹 : 2𝐴𝑅 → 2𝐴𝑅 and
𝐹𝐴𝐹 (𝑆) = {𝐴|𝐴 is acceptable with respect to 𝑆}.
(Grounded Extension) The grounded extension of an argumentation framework AF, denoted by
𝐺𝐸𝐴𝐹 , is the least fixed point of 𝐹𝐴𝐹 with respect to set-inclusion [13]. Since 𝐹𝐴𝐹 is a monotonic
function with respect to set inclusion [13], the existence of the fixed point for this function
follows from the Knaster-Tarski theorem [25].
It can be proved that the grounded extension of the abstract argumentation framework
utilized in the proposed argumentation-based learning method is the singleton admissible sets
which do not have both incoming and outgoing edges.
2.2. Formal Definition of an Abstract Bipolar Argumentation Framework
An Abstract Bipolar Argumentation Framework (BAF ) [14] is an extension of Abstract Ar-
gumentation Framework by adding a support relationship. A BAF is a triple of the form
< 𝐴𝑅, 𝑅att , 𝑅sup > where AR is the finite set of arguments, 𝑅att ⊆ 𝐴𝑅 × 𝐴𝑅 is the attack set and
𝑅sup ⊆ 𝐴𝑅 × 𝐴𝑅 is the support set. Considering 𝐴𝑖 and 𝐴𝑗 ∈ 𝐴𝑅, then 𝐴𝑖 𝑅att 𝐴𝑗 means that 𝐴𝑖
attacks 𝐴𝑗 and 𝐴𝑖 𝑅sup 𝐴𝑗 means that 𝐴𝑖 supports the argument 𝐴𝑗 .
The semantics of BAF are as follows:
(Conflict-Free) Let 𝑆 ⊆ 𝐴𝑅. S is conflict-free iff there is no 𝐵, 𝐶 ∈ 𝑆 such that B attacks C.
(Admissible set) Let 𝑆 ⊆ 𝐴𝑅. 𝑆 is admissible iff 𝑆 is conflict-free, closed for 𝑅sup (if 𝐵 ∈ 𝑆 and
𝐵 𝑅𝑠𝑢𝑝 𝐶 ⇒ 𝐶 ∈ 𝑆) and 𝑆 defends all its elements.
(Preferred extension) The set 𝐸 ⊆ 𝐴𝑅 is a preferred extension iff 𝐸 is inclusion-maximal
among the admissible sets. An inclusion-maximal set among a collection of sets is a set that is
not a subset of any other set in that collection.
(Supporting Weights) Like [26] the support relations in our model also have an assigned
weight. Therefore, a node with higher sum of supporting weights can attack nodes with lower
sum of supporting weights.
2.3. Formal Definition of Online Incremental Machine Learning Algorithms
We define an incremental learning approach that uses a sequence of data instances 𝑑1 , 𝑑2 , ..., 𝑑𝑡
for generating the corresponding models 𝑀1 , 𝑀2 , ..., 𝑀𝑡 . In case of incremental online learning,
each data instance 𝑑𝑖 incrementally updates the model and 𝑀𝑖 ∶ ℝ𝑛 → {1, ..., 𝐶}, where 𝐶 is the
number of class labels, is representing the model which depends on 𝑀𝑖−1 . The online learning
is then defined as an incremental learning which is also able to continuously learn. Incremental
learning approaches have the following properties:
• The model should adapt gradually, i.e. 𝑀𝑖 is updated using 𝑀𝑖−1 .
• The previously learned knowledge should be preserved.
A recent study on the comparison of the state-of-the-art methods for incremental online machine
learning [27] shows that Incremental Support Vector Machines (ISVM) [28, 29] together with
LASVM [30], which is an online approximate SVM solver, and Online Random Forest (ORF )
[31] outperform the other methods. The comparison methods used in our paper have been
chosen based on the aforementioned survey [27].
The proposed argumentation-based incremental learning approach uses the bipolar argu-
mentation framework to model the visited data instances and generate relevant hypotheses.
Subsequently, the abstract argumentation framework is used to model the defeasibility relations
(i.e. the attack relations) between the current set of generated hypotheses and predict the best
action (recovery behavior) for an unforeseen incoming data instance. Furthermore, the model
incrementally gets updated as new data instances enter the model.
3. Scenarios
The performance of the different methods is tested using two test scenarios. The aim of the
first test scenario is to model a situation where a programmer has provided an initial solution
(e.g., a top level behavior such as entering the room), while (s)he has not accounted for all
possible failures (e.g., objects and persons blocking the entrance), but allows the robot to find
new solutions whenever a (previously unseen) failure occurs.
The basic setup of the first test scenario is illustrated in Fig. 1. The high-level behavior of
the robot aims to proceed from the initial location to the target location using three entrances.
Different obstacles might be on its way to the target location. In these scenarios, an agent
observes all the obstacle locations at once and chooses a single recovery behavior (action) for
recovering from that failure state. The agent can reach the goal if it chooses the best recovery
behavior; otherwise, it fails to reach the goal.
3.1. Recovery Behaviors
Whenever the robot is confronted with a failure state, it may use any of the following recovery
behaviors to resolve the issue. The run-time of each recovery behavior in seconds is presented
1 2 3
Target Location
4 5 6
Alt. Alt.
7 8 9
Initial Location
Figure 1: Schematic overview of the possible failure state scenario. Only the green locations are relevant
for finding the best recovery behavior. Alt. stands for the Alternative Route recovery behavior.
in parentheses in front of each recovery behavior:
• Continue (2s): This solution is only useful if the failure has resolved itself (e.g., the obstacle
moved away just after the failure).
• Push (5s): The robot can try pushing any obstacle.
• Ask (4s): The robot can try to ask any type to move.
• Alternative Route (Alt) (10s): The robot can move to another entrance to reach the target
location.
It is important to note that choosing Alternative Route as the best recovery behavior may not
always lead to success, because the robot may again be confronted with new obstacles (Fig.
1). Moreover, the best recovery behavior not only depends on the run-time of each recovery
behavior, but also on the type, the color and the location of the obstacles.
3.2. Test Scenario 1
In this scenario, three types of obstacles (ball, box or person) with four colors (red, blue, green
or yellow) can be presented in one of the locations 1 to 9 (Fig. 1). There can be either zero
or one combination of color-type in each location. Only location number 5 and 8, marked in
green (Fig. 1), is relevant for choosing the best recovery behavior. It is important to notice
that the robot does not know this fact and it should infer that the only effective locations are
location number 5 and 8 by observing different failure states in the environment. The agent
observes all the obstacle locations at once and chooses a single recovery behavior (action) at
each state. A new state is generated randomly at each time step. The number of possible
combinations of the color-type in each location is 13 (3 types × 4 colors + “no obstacle” = 13).
Since there are 9 locations in this scenario, the number of all possible states in this scenario is
139 = 10, 604, 499, 373.
3.3. Test Scenario 2
The second scenario has a different purpose and context. It shows the applicability of the
proposed method outside the robotics field. The recent study on online incremental machine
learning techniques [27] used the publicly available datasets from the UCI machine learning
repository [32]. We also used the SPECT heart dataset from the UCI machine learning repository.
4. Method
In this section, we will discuss the proposed argumentation-based learning method for recover-
ing from an unforeseen failure state1 .
4.1. Argumentation-Based Learning (ABL)
In order to explain ABL, we first use a simplified version of the previous test scenarios where
there is only one location ahead of the robot (instead 9). When there is no obstacle ahead of the
robot, the best recovery behavior is “Continue”.
Assume that the robot confronts a blue-ball blocking the entrance. Since there is no pre-
trained model yet, the robot tests different recovery behaviors in order of their run-time to find
the best one. Supposing that pushing the ball was successful in this case, the robot should learn
from this experience.
However, unlike the traditional tabular reinforcement learning techniques, only learning the
best recovery behaviors (actions) for exactly the same experiences (states) is not enough. We
need a learning approach capable of inferring the correlated feature values (each feature value is
the color or type of the obstacle at each location or an empty location with no color and type) for
choosing the best recovery behavior. This is known as generalization in the machine learning
literature. For instance, confronting a red ball and a green ball with the same recovery behavior
of pushing, the robot should make a new hypothesis push a ball. Therefore, the next time the
robot confronts the yellow ball, it can easily infer that Push is the best recovery behavior.
Confronting a yellow ball with Alternative Route as the best recovery behavior contradicts
the previous hypothesis. Therefore, a new hypothesis is made: Push a ball unless it’s yellow.
From an argumentation perspective, we can see each hypothesis as an argument. Therefore,
the second generated hypothesis can attack and defeat the first argument. This is inspired by
human agents who make new hypotheses from their perceptions and reason about the best
course of action at each state.
The architecture of the proposed argumentation-based learning method is shown in Fig.
2. A bipolar argumentation framework is used as hypotheses generator unit and an abstract
argumentation framework models the defeasibility relation between these generated hypotheses.
When a new data instance enters the model, all the combinations of its feature-values and
the set of nodes in the grounded extension of the 𝐴𝐹 will be extracted. Each node (argument)
in the AF unit is of the form precondition → post-condition: weight. According to the similarity
between the preconditions of the arguments in the grounded extension and the feature values
combinations, there will be three possible cases. Either there will be a unique similarity,
1
See [1] for a more detailed and a more formal definition of the proposed method.
Figure 2: Architecture of the proposed Argumentation-based learning method.
multiple similarities or no similarity. In case of unique similarity, the post-condition of the
argument (which is a recovery behavior) will be used as the first guess and will be applied
to the environment to see the result. On the occasion that there exist multiple similarities,
the recovery behavior with the highest weight among the arguments will be chosen and its
post-condition will be applied to the environment. A successful recovery from the failure state
will update the 𝐵𝐴𝐹 unit. On the other hand, failure from recovery will lead to generating the
second guess, updating the 𝐵𝐴𝐹 unit, generating hypotheses from 𝐵𝐴𝐹 unit and updating the
𝐴𝐹 unit, respectively.
We now use an illustrative example to explain the proposed method in more detail.
4.2. Example
Table 1 shows the best recovery behavior when the robot confronts an obstacle with different
colors and types. Notice that this table is only used for this example and a randomly generated
table is utilized for each of the 1000 independent runs for the experiments. Figure 3 shows the
updating procedure of the model step by step. In the hypotheses generation unit (BAF ), an
arrow → shows a support relation between arguments and ↛ shows an attack relation between
them. However, in AF, → shows an attack relationship between the arguments.
Referring to Table 1, at the beginning of the learning procedure, the robot confronts a Red-Ball
(R-Ba). It tests all the recovery behaviors in order of their run-times and finds the Push recovery
behavior as a success (Table 1). Subsequently, the Bipolar Argumentation Framework is getting
updated as in Fig. 3. In order to update the BAF, first, the best recovery node is added which
is Push in this case. Then all the possible combinations of the feature-values of the current
state are added as supporting nodes. The supporting nodes for Push are R, Ba and R-Ba. If
there previously exists the same supporting node, its supporting weight will be increased. For
instance in Fig. 3, where 8:B-Bo enters the BAF, since B and B-Bo are new supporting nodes for
the Alt (Alternative Route) recovery behavior, they are added to the model with a supporting
Order Color Type Best Recovery Behavior
1 Red Ball Push
2 Red Box Alternative Route
3 Red Person Ask
4 Green Ball Push
5 Green Box Alternative Route
6 Green Person Ask
7 Blue Ball Push
8 Blue Box Alternative Route
9 Blue Person Alternative Route
10 Yellow Ball Push
11 Yellow Box Alternative Route
12 Yellow Person Ask
13 None None Continue
Table 1
Possible combinations of color-type with the best recovery behaviors.
weight equal to 1. On the other hand, Bo already exists in the set of supporting nodes for Alt and
its weight is increased. After updating the supporting weights, a set of hypotheses is generated
based on the number of occurrences of each supporting node. For instance, after observing
1:Red-Ball (R-Ba), R → Push and Ba → Push are added to the AF unit.
Confronting 2:R-Bo and using the previously generated hypotheses (specifically R → Push),
the robot would infer that the best possible recovery behavior is Push, which is a wrong choice
in this case (Table 1). Therefore, the robot tries other recovery behaviors and finds Alt as success
and updates the model accordingly. Moreover, a bidirectional attack will be added among all the
recovery nodes in the BAF (in this case, Alt and Push). Subsequently, the new set of hypotheses
is generated to update the hypotheses argumentation unit. Finally, an abstract argumentation
framework is updated to model the attack relations between the set of generated hypotheses
(arguments). This BAF-AF update cycle goes on and on during the learning procedure.
In this small example, seven out of thirteen predictions of the model are correct, and only
two are wrongly classified using the proposed argumentation-based learning. In other cases,
our system can provide multiple probable guesses. For instance, when 12:Y-P enters the system
in Fig. 3, the AF cannot provide any suggestion but the BAF will suggest both Ask and Alt
as the candidate recovery behaviors. However, the mapping of the states to the best recovery
behavior is randomly generated in all the experiments.
4.3. Hypotheses Generation Unit (BAF Unit)
This unit has two roles. Firstly, it generates a new set of hypotheses whenever the AF unit
could not classify the new data instance correctly (1). The second role of this unit is to produce
a second guess for the best recovery behavior (2):
1) In order to generate a new set of hypotheses from the constructed BAF, only one recovery
behavior is considered which is highlighted with a red box in Fig. 3.
The only nodes which are getting updated during this process are the best recovery behavior
Figure 3: Example of Argumentation-Based Learning for the example scenario. Here only observations
number 1, 2, 3, 9, 10 and 12 of Table 1 are shown selectively.
Figure 4: The generated BAF when Yellow-Person (12:Y-P ) enters the model. Blue nodes show the
intersection of preferred extensions and recovery behavior nodes.
for the current data instance and its supporting nodes. Autonomously identifying the best
recovery behavior through trial and error, the update procedure for hypotheses generation takes
place. The updating procedure searches for a node in the BAF graph with the best recovery
behavior and appends all the possible combinations of the feature-values of the current state to
the support nodes of the best recovery behavior node. In case that a supporting node already
exists in the best recovery behavior node, its supporting weight is incremented.
2) In order to generate a second guess, a new BAF should be constructed. For an unfore-
seen failure state, the set of all possible combinations of feature-values is compared with the
supporting nodes of each recovery behavior node. According to the sum of the matching
supporting weights, the attack relations are adapted among the recovery behaviors. Therefore,
only recovery behaviors with a higher sum of the matching supporting weights can attack
the other recovery behavior. For instance, in the example, when 12: Y-P enters the model for
prediction, the AF is not be able to guess the best recovery behavior. Constructing a new BAF
for a second guess, shown in Fig. 4, the calculated weighted sum for the Alternative Route (Alt)
node is the same as Ask and higher than Push. Accordingly, the attack relations get updated.
Using preferred extension semantics and its intersection with recovery behavior nodes, both
Alternative Route (Alt) and Ask are chosen as the second guesses.
4.4. Hypotheses Argumentation Unit using AF
As stated in the previous sections, this unit tries to justify what has been learned so far by
updating the attack relations between the arguments (hypotheses). The arguments in this
framework can only bidirectionally attack each other when they have the same preconditions
but different post-conditions.
When a new data instance enters the model, there are three possible cases for the set of
hypotheses in the grounded extension of the AF. When the grounded extension of the AF
is the empty set, the second guess is generated by the BAF unit. If one argument with the
same post-condition exits in the grounded extension of the AF, then this post-condition will
be the AF ’s first guess. If more than one argument with different recovery behaviors in their
post-condition was chosen, the weights of arguments determine which argument has more
power to be selected. For instance in the example, if blue-ball enters the model after it has
been trained using the complete set of data in Table-1, both B → Alt: 2/4 and Ba → Push:1 can
be used for prediction. Since the Ba → Push:1 has higher weight, the Push recovery behavior
will be chosen, which is the correct choice for this failure state. Notice that in the proposed
argumentation-based learning method, it can be proved that the grounded extension is a set of
the singletons in the AF.
5. Experiments
In this section, we compare the performance of our proposed ABL method with other incremental
learning techniques and an approximation-based reinforcement learning algorithm. The survey
by V. Losing et al. compared a broad range of incremental online machine learning techniques
[27]. Using the key methods in their survey, we are also comparing the proposed method with
Incremental Support Vector Machine (ISVM) [28, 33, 34], incremental decision tree based on
C4.5 [35] and ID3, incremental Bayesian classifier [36], Online Random Forest (ORF )[31] and
Multi-Layer Neural Networks for classification with localist models like Radial Basis Functions
(RBF ) which work reliably in incremental settings [37, 38].
5.1. Comparison criteria
In the robotic scenario, we need a learning approach which can quickly learn to recover from
failure states in a low number of attempts. Moreover, for the other test scenario, the goal is
to incrementally learn from a lower number of training instances. Therefore, the increase in
learning precision in a lower number of attempts is one important criterion (which we call
learning speed) to evaluate the efficiency of the method [39]. Therefore, learning curves with
the highest steepness in a smaller number of attempts are desirable. Furthermore, the final
learning precision is also an important criterion.
5.2. Results
As one can see in Fig. 5a and Fig. 5b, the proposed Argumentation-Based Learning (ABL)
method outperforms all the other methods in both the comparison criteria used for this research,
namely, the final learning precision and the learning speed. The steepness of the learning curve
shows that the ABL learns faster in a lower number of iterations.
For the first test scenario, after observing 30 failure states, ABL achieves 74% precision, while
the best method among others has 60% precision. The final precision of ABL is 95%, while the
best final precision among other methods is 90%.
In the second scenario, which differs from the prior scenario in context, ABL repeatedly out-
performs all the other methods in both of the comparison criteria. Among other methods,
incremental naive Bayes and incremental random forest (ORF ) have better results. The final
learning precision of ABL in this scenario is 75% while it is 70% for the incremental naive Bayes
(a) First Scenario. (b) Second Scenario
Figure 5: Comparison of Argumentation-Based Learning (ABL) with key methods for incremental
online learning [27] using a) the first test scenario b) the second test Scenario.
method. The slope of the learning curve also shows the faster learning speed of ABL with
respect to all of the other methods.
6. Discussion
A key reason that the proposed method works better than Naive Bayes originates from the
independence assumption between all features in the Naive Bayesian formulation. In the case
of neural networks, considering that there is only a small number of training data instances,
a complex neural network tends to over-fit and a small neural network leads to under-fitting.
Choosing the best neural network architecture dynamically according to the number of visited
data is also a challenging task. On the other hand, decision-tree based techniques fail at the
initial recovery attempts and then gradually learn the best recovery behavior. This is because of
the change in entropy or information gain when new unforeseen data updates the decision tree.
This is also the case with the Online Random Forest (ORF ) method. Furthermore, ISVM does
not perform well in circumstances where only a few features are associated with predicting the
class label. In all the above cases, the suggested ABL approach performed better as it considers
any possible dependence between features and it can immediately focus on features which are
most relevant for the optimal decision.
Moreover, ABL leads to an explicit representation of the learning process understandable for
humans, as is also the case with decision-tree based techniques. In contrast, neural networks,
support vector machines and Bayesian techniques are all black boxes [40] (this means that the
trained models are not easily interpretable and explainable) for the humans. This explicit repre-
sentation of the learning process can be utilized in combination with human-robot interaction.
Employing this property, ABL can be used in multi-agent scenarios where agents can transfer
their knowledge to each other.
Consequently, the proposed argumentation-based incremental learning algorithm could learn
in fewer attempts with higher precision than other algorithms used for comparison. Moreover,
ABL extracts an explicit set of rules that explain the knowledge acquired by the agent over the
interaction with the environment. This feature makes the method more explainable and easy to
debug by an expert.
Therefore, this method can be a good alternative when the feature values are discrete. Al-
though we have shown that the current ABL approach is working well for the aforementioned
scenarios in this paper, these results are limited to datasets with discrete feature values that are
not high-dimensional. To make ABL more efficient for higher dimensional problems, we have
introduced Accelerated Argumentation-Based Learning (AABL) [41] to improve the space and
computational complexity of the method.
7. Conclusion
General purpose service robots should be able to recover from unexpected failure states caused
by environmental changes. In this article, an argumentation-based learning (ABL) approach
is proposed which is capable of generating relevant hypotheses for online incremental learn-
ing scenarios. This set of hypotheses is updated incrementally when unforeseen data enters
the model. The conflicts among these hypotheses are modeled by Abstract Argumentation
Frameworks.
The performance of ABL has been evaluated using both the robotics and the non-robotics
incremental learning scenarios. The second scenario, which has a non-robotic context, is a
publicly accessible dataset from the UCI machine learning repository. This scenario shows the
fact that the proposed ABL method can be used in any online incremental learning application
with discrete feature values. According to these experiments, the proposed method learns
faster and with higher ultimate classification precision than various state-of-the-art online
incremental learning methods.
References
[1] H. Ayoobi, M. Cao, R. Verbrugge, B. Verheij, Argumentation-based online incremental
learning, IEEE Transactions on Automation Science and Engineering (2021) 1–15. doi:1 0 .
1109/TASE.2021.3120837.
[2] V. N. Lu, J. Wirtz, W. H. Kunz, S. Paluch, T. Gruber, A. Martins, P. G. Patterson, Service
robots, customers and service employees: what can we learn from the academic literature
and where are the gaps?, Journal of Service Theory and Practice (2020).
[3] M. Mende, M. L. Scott, J. van Doorn, D. Grewal, I. Shanks, Service robots rising: How hu-
manoid robots influence service experiences and elicit compensatory consumer responses,
Journal of Marketing Research 56 (2019) 535–556.
[4] S. Schneider, F. Hegger, A. Ahmad, I. Awaad, F. Amigoni, J. Berghofer, R. Bischoff, A. Bonar-
ini, R. Dwiputra, G. Fontana, et al., The RoCKIn@Home Challenge, in: ISR/Robotik 2014;
41st International Symposium on Robotics, 2014, pp. 1–7.
[5] Y. Jiang, N. Walker, J. Hart, P. Stone, Open-world reasoning for service robots, in:
Proceedings of the International Conference on Automated Planning and Scheduling,
volume 29, 2019, pp. 725–733.
[6] A. Kuestenmacher, N. Akhtar, P. G. Plöger, G. Lakemeyer, Towards robust task execution
for domestic service robots, Journal of Intelligent & Robotic Systems 76 (2014) 5–33.
[7] K. Talamadupula, G. Briggs, M. Scheutz, S. Kambhampti, Architectural mechanisms
for handling human instructions for open-world mixed-initiative team tasks and goals,
Advances in Cognitive Systems 5 (2017) 37–56.
[8] P. Schermerhorn, M. Scheutz, Using logic to handle conflicts between system, component,
and infrastructure goals in complex robotic architectures, in: 2010 IEEE International Con-
ference on Robotics and Automation, 2010, pp. 392–397. doi:1 0 . 1 1 0 9 / R O B O T . 2 0 1 0 . 5 5 0 9 7 6 3 .
[9] F. H. Van Eemeren, B. Garssen, E. C. Krabbe, A. F. S. Henkemans, B. Verheij, J. H. Wagemans,
Handbook of Argumentation Theory, Dordrecht: Springer, 2014.
[10] L. Rizzo, L. Longo, An empirical evaluation of the inferential capacity of defeasible
argumentation, non-monotonic fuzzy reasoning and expert systems, Expert Systems with
Applications 147 (2020) 113220. URL: https://www.sciencedirect.com/science/article/pii/
S0957417420300464. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . e s w a . 2 0 2 0 . 1 1 3 2 2 0 .
[11] A. Vassiliades, N. Bassiliades, T. Patkos, Argumentation and explainable artificial intelli-
gence: a survey, The Knowledge Engineering Review 36 (2021).
[12] K. Atkinson, T. Bench-Capon, D. Bollegala, Explanation in AI and law: Past, present
and future, Artificial Intelligence 289 (2020) 103387. URL: https://www.sciencedirect.
com/science/article/pii/S0004370220301375. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . a r t i n t . 2 0 2 0 .
103387.
[13] P. M. Dung, On the acceptability of arguments and its fundamental role in nonmono-
tonic reasoning, logic programming and n-person games, Artificial Intelligence 77 (1995)
321–357.
[14] L. Amgoud, C. Cayrol, M.-C. Lagasquie-Schiex, P. Livet, On bipolarity in argumentation
frameworks, International Journal of Intelligent Systems 23 (2008) 1062–1093.
[15] O. Cocarascu, F. Toni, Argumentation for Machine Learning: A Survey, in: COMMA, 2016,
pp. 219–230.
[16] M. Mozina, J. Zabkar, I. Bratko, Argument based machine learning, Artificial Intelligence
171 (2007) 922–937.
[17] P. Clark, T. Niblett, The CN2 Induction Algorithm, Machine Learning 3 (1989) 261–283.
[18] L. Amgoud, M. Serrurier, Agents that argue and explain classifications, Autonomous
Agents and Multi-Agent Systems 16 (2008) 187–209.
[19] L. Carstens, F. Toni, Using argumentation to improve classification in natural language
problems, ACM Transactions on Internet Technology (TOIT) 17 (2017) 30.
[20] N. Kotonya, F. Toni, Gradual argumentation evaluation for stance aggregation in au-
tomated fake news detection, in: Proceedings of the 6th Workshop on Argument Min-
ing, Association for Computational Linguistics, Florence, Italy, 2019, pp. 156–166. URL:
https://aclanthology.org/W19-4518. doi:1 0 . 1 8 6 5 3 / v 1 / W 1 9 - 4 5 1 8 .
[21] M. Moens, Argumentation mining: How can a machine acquire common sense and world
knowledge?, Argument and Computation 9 (2018) 1–14. doi:1 0 . 3 2 3 3 / A A C - 1 7 0 0 2 5 .
[22] S. Eger, J. Daxenberger, I. Gurevych, Neural end-to-end learning for computational
argumentation mining, in: Proceedings of the 55th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), Association for Computational
Linguistics, Vancouver, Canada, 2017, pp. 11–22. URL: https://www.aclweb.org/anthology/
P17-1002. doi:1 0 . 1 8 6 5 3 / v 1 / P 1 7 - 1 0 0 2 .
[23] M. Bishop, C. Gates, K. Levitt, Augmenting machine learning with argumentation, in:
Proceedings of the New Security Paradigms Workshop, NSPW ’18, ACM, New York,
NY, USA, 2018, pp. 1–11. URL: http://doi.acm.org/10.1145/3285002.3285005. doi:1 0 . 1 1 4 5 /
3285002.3285005.
[24] H. Ayoobi, M. Cao, R. Verbrugge, B. Verheij, Local-HDP: Interactive Open-ended 3D Object
Category Recognition in Real-Time Robotic Scenarios, Robotics and Autonomous Systems
(2021).
[25] A. Tarski, A lattice-theoretical fixpoint theorem and its applications., Pacific Journal of
Mathematics 5 (1955) 285–309.
[26] A. Pazienza, S. Ferilli, F. Esposito, On the gradual acceptability of arguments in bipolar
weighted argumentation frameworks with degrees of trust, in: International Symposium
on Methodologies for Intelligent Systems, Springer, 2017, pp. 195–204.
[27] V. Losing, B. Hammer, H. Wersing, Incremental on-line learning: A review and comparison
of state of the art algorithms, Neurocomputing 275 (2018) 1261–1274.
[28] G. Cauwenberghs, T. Poggio, Incremental and decremental support vector machine
learning, in: Advances in Neural Information Processing Systems, 2001, pp. 409–415.
[29] A. Soula, K. Tbarki, R. Ksantini, S. B. Said, Z. Lachiri, A novel incremental Ker-
nel Nonparametric SVM model (iKN-SVM) for data classification: An application to
face detection, Engineering Applications of Artificial Intelligence 89 (2020) 103468.
URL: https://www.sciencedirect.com/science/article/pii/S0952197619303501. doi:h t t p s :
//doi.org/10.1016/j.engappai.2019.103468.
[30] A. Bordes, S. Ertekin, J. Weston, L. Bottou, Fast kernel classifiers with online and active
learning, Journal of Machine Learning Research 6 (2005) 1579–1619.
[31] A. Saffari, C. Leistner, J. Santner, M. Godec, H. Bischof, On-line random forests, in:
Computer Vision Workshops (ICCV Workshops), 2009 IEEE 12th International Conference
on, IEEE, 2009, pp. 1393–1400.
[32] D. Dua, C. Graff, UCI machine learning repository, 2017. URL: http://archive.ics.uci.edu/ml.
[33] B. Gu, V. S. Sheng, K. Y. Tay, W. Romano, S. Li, Incremental support vector learning for
ordinal regression, IEEE Transactions on Neural Networks and Learning Systems 7 (2015)
1403–1416.
[34] B. Gu, V. S. Sheng, Z. Wang, D. Ho, S. Osman, S. Li, Incremental learning for 𝜈-support
vector regression, Neural Networks 67 (2015) 140–150.
[35] J. R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Publishers Inc.,
San Francisco, CA, USA, 1993.
[36] R. Agrawal, R. Bala, Incremental Bayesian classification for multivariate normal distribu-
tion data, Pattern Recognition Letters 29 (2008) 1873 – 1876. URL: http://www.sciencedirect.
com/science/article/pii/S0167865508001992. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . p a t r e c . 2 0 0 8 .
06.010.
[37] P. Reiner, B. M. Wilamowski, Efficient incremental construction of RBF networks using
quasi-gradient method, Neurocomputing 150 (2015) 349–356.
[38] J. Lu, F. Shen, J. Zhao, Using Self-Organizing Incremental Neural Network (SOINN) for
radial basis function networks, in: 2014 International Joint Conference on Neural Networks
(IJCNN), IEEE, 2014, pp. 2142–2148.
[39] H. Ayoobi, M. Rezaeian, Swift distance transformed belief propagation using a novel
dynamic label pruning method, IET Image Processing 14 (2020) 1822–1831.
[40] B. Zhou, D. Bau, A. Oliva, A. Torralba, Interpreting deep visual representations via network
dissection, IEEE Transactions on Pattern Analysis and Machine Intelligence 41 (2019)
2131–2145. doi:1 0 . 1 1 0 9 / T P A M I . 2 0 1 8 . 2 8 5 8 7 5 9 .
[41] H. Ayoobi, M. Cao, R. Verbrugge, B. Verheij, Argue to learn: Accelerated argumentation-
based learning, in: 20th IEEE International Conference on Machine Learning and Applica-
tions (ICMLA), IEEE, 2021.