Autonomía del Delito de Lavado de Activos y Prueba Indiciaria

Reviewing the literature reveals that most published approaches focus either on establishing astudent modelfilled with knowledge about the learner’s skill mastery and/or engagement, or on the develop- ment of sophisticated planning algorithms to choose the next steps to be taken in the tutoring inter- action (pedagogical module). However, the work on ITSs suggests that both, a sophisticatedstudent modelandpedagogical module, are equally essential for building a tutoring system to provide lessons as beneficial as possible. Because of this, some planning approaches already try to combine both, e.g., by classifying children’s learner type to provide well designed default learning content suitable for the re- spective needs and preferences (e.g.,Hatzilygeroudis and Prentzas,2004;Castellano et al.,2007). But although these models are already rather complex in their planning mechanics, e.g., by basing it on neural networks or even on a hybrid approach that merges symbolic rules and neurocomputing, they just aim for sorting the learner into rather rough groups instead of keeping track of her knowledge to establish an elaboratedstudent model.

Also first approaches based on RL can be found that incorporate a rudimentarystudent model(Mal- pani et al.,2011) or a small history of the learner’s answers (Sarma and Ravindran,2007) to represent or infer her already attained knowledge. In fact, they demonstrated that such fairly simple extensions can already increase the efficiency of an ITS. Moreover, basing the planning on RL allows for learning the pedagogical rules from a dataset or the interaction itself, which was shown to yield good results (Chi et al.,2009). However, these approaches to implement thepedagogical modulecontain also some weaknesses. The resulting models are often inflexible, because they are designed for a specific type of task, e.g., a sequential decision task (e.g.,Cakmak and Lopes,2012), tested only on simulations without the unpredictable noise of real learners (e.g.,Sarma and Ravindran,2007;Malpani et al.,2011) or ignore the uncertainty about learners’ real knowledge (e.g.,Cakmak and Lopes,2012). In addition, employing RL for larger learning domains, e.g., language learning, or extending thestudent modelby considering also learners’ engagement, i.e., affective and cognitive states, enlarges the state space and, thus, strongly increases the time needed to personalize and to learn how to behave until it can be used in the wild effectively. Nevertheless, all these models already demonstrated that basing the planning mechanics of thepedagogical moduleon knowledge about the student, even when this information base is fairly simple, can significantly increase the effectiveness of an ITS.

In general, to establish a more elaboratedstudent modeltwo different types of approaches for track- ing learners’ knowledge are commonly used. First, assumption-free approaches that allow for the model to learn the latent structures within a dataset by itself and, thus, make the time-consuming task of defining them by hand unnecessary. But this increases the difficulty of keeping track of what the system learned and how it will behave in each situation, which, however, is precisely what is often de- sired in the realm of developing educational software. Experts want to control what and how children should learn, so that it is in line with research in pedagogic and pedagogical psychology. Furthermore, there is a lack of huge datasets to train these models. Consequently, this type of model does not provide the best basis for the development of a SARTS.

5.1. Model Selection

Assumption-based models, instead, require predefined assumptions, preferably provided by hu- man experts, which complicates their development, but often maintain their interpretability and un- derstandability. Commonly used assumption-based knowledge tracing approaches build on Bayesian methods, but the questions about the optimal choice between those and logistic models is not fully resolved yet. Although these models already have been compared (Gong et al.,2010), the results are not conclusive, probably do not generalize properly and, thus, are not applicable to each domain. In general, logistic models are favored for memory building processes since the knowledge state is mod- eled more naturally by gradual changes. In contrast, Bayesian methods, e.g., BKT, often apply a more discrete state transition from not known to known, which allows to model the understanding and sense-making processes in fine-grained knowledge components. Furthermore, the latter often provides a specific graphical structure to model influences of each included attribute. This enables one to under- stand the likely effects and decisions of such models easily and, thus, allows to prove their validity in the educational context. Additionally, a subset of these approaches allow for a simulation of the learner’s development based on the influences of different variables to plan the next tutoring steps. They further seem to be easily extendable and, thus, provide a suitable basis for an incremental implementation of a student model. In fact,Khajah et al.(2014a) showed that a BKT can easily be transformed into a hybrid model by including the problem difficulty or general student abilities as latent factors (based on logistic models) directly into the BKT so that they can influence the tracing process explicitly. Their results showed that a BKT model with these rather simple extensions can already outperform the traditional BKT, as well as basic logistic models (Khajah et al.,2014a).

In addition, since BKT is a specific type of DBN, it can easily be extended by a decision component, e.g., to choose the next problem difficulty, leading to a model similar to a Dynamic Bayesian Deci- sion Network (DBDN) (Russell and Norvig,2010, p. 664ff.) or a POMDP (Russell and Norvig,2010, p. 658ff.). While solutions of the latter are optimal plans that are computed offline and conditioned on future observations, the former can be regarded as a computational representation of POMDPs, which determines solutions for finite time horizons in an online fashion (cf.Polich and Gmytrasiewicz,2007). This enables the ITS to select actions and to reason about their effects, so that the learners will receive a tutoring interaction which they most likely benefit from. However, finding an optimal action pol- icy is often computationally intractable. Although research on compressing the state space by means of new representation methods to make them tractable even in larger learning domains has already been done (Folsom-Kovarik et al.,2013), this approach might not be applicable in all domains and, in particular, not in language learning. Another possibility is to use online-planning algorithms, e.g., a heuristic forward search (Brunskill and Russell,2011;Rafferty et al.,2011). However, this type of algo- rithm is not guaranteed to explore the whole state and action space to find the optimal solution, since in larger spaces it is limited through a time constraint of just a few seconds to maintain its tractability. But nevertheless, within a domain with a limited or smartly defined state space and a suitable online planning algorithm, POMDPs or DBDNs, respectively, still provide a huge potential to combine in- formation stored in thestudent modelwith thepedagogical modulefor planning the next steps in a tutoring interaction to optimally address the learners’ cognitive learning.

In conclusion, extending the traditional BKT to build a DBDN poses a promising basis for imple- menting and combining thestudent modelandpedagogical moduleof an SARTS. First, the underlying BKT is easily extendable and, hence, can also handle further information about the learner, e.g., the engagement. Second, integrating the decision-making right into the BKT allows for actions to influ- ence the tracing process, which, in fact, was already shown to improve the tracing results (Khajah et al.,

2014a). Third, in the chosen domain of word learning of a foreign language, the state and action spaces can be modeled so that the planning process stays tractable. In general, the state portrays the words to be learned whose number, however, can be several tens of thousands. But, the state space can be divided into small “chunks” so that each chunk just includes a small portion of the vocabulary. Let- ting the planning algorithm just work within the current chunk reduces the costs for planning the next steps dramatically. After a chunk is mastered by the student, the system simply switches to the next chunk of skills and goes on with the teaching process. This is also an established practice in traditional classroom environments, in which teachers also do not teach the whole vocabulary of a language at a time. In addition to the already mentioned benefits, modeling thestudent modelandpedagogical mod- uletightly coupled also allows to represent some of the interconnections between the different learning dimensions. This is important since the inevitable influences of other dimensions can change the profit for the actions executed by a SARTS (see Section2.1.6). However, incorporating the influences right into the model enables the tutoring system to simulate different actions and action combinations based on the current information about the learner and to choose them accordingly to optimize the tutoring experience with respect to all dimensions of learning.