Human-Aware Planning for Promoting Social Behavior-Change in Autism
An action reasoning approach utilizing answer set programming
Andreas Br¨annstr¨om
Andreas Br ¨annstr ¨om Spring 2020
Master’s Thesis in Computing Science, 30 credits
Supervisors: Juan Carlos Nieves Sanchez, Timotheus Kampik Examiner: Jerry Eriksson
Master Programme in Computing Science, 120 credits
This study aims to find ways for an intelligent software agent to un- derstand a human’s behavior in a social situation, and in this process, plan its own actions in order to assist the human in reaching their goals.
The use-case is a Virtual Reality (VR) game, developed for children with autism for practicing social scenarios, scenarios that children with autism often find stressful or scary. This study investigates how an in- telligent software agent can, by continuously analyzing how the user interacts with a simulated social scenario, adapt the simulation through appropriate interventions. This helps the user to succeed in the social scenario while providing a challenging learning environment. In this human-aware planning problem, the variables of the environment and the human’s mental state constitutes Interaction Constraints (IC) for the system. Central questions in this study regard what, how and when ap- propriate interventions can be provided by the system to facilitate be- havior change, and in this process, preserve dynamic sub-goals of the human. An action reasoning computational model is proposed inspired by three cognitive theories; (1) the theory of planned behavior, answer- ing the what and how questions of the model by defining transitions between goals; (2) the stress staircase and (3) the zone of proximal de- velopment, together answering the when questions of the model. By defining the user’s physical and mental state based on fluents of the en- vironment, plans can be generated for providing goal-oriented interven- tions. A figurative instance of the scenario can then evaluate the plans, according to weights tailored to the individual user, to select a final plan for adapting the virtual scenario. The proposed human-aware planning architecture can also be applied in environments that are not virtual, by utilizing modern mobile devices which have built-in sensors that mea- sure motion, orientation, and various environmental conditions. Future work concerns how automated learning approaches can be incorporated in the architecture to provide tailored levels of personalization.
I would like to thank my supervisors Juan Carlos Nieves Sanchez and Timotheus Kampik for all the guidance and valuable insights throughout this project. Your support greatly improved the quality of this work. I want to thank all the specialists for their participation in the study. Your knowledge comprises the backbone of this thesis.
1 Introduction 1
1.1 Human-aware planning 1
1.2 The aim of this thesis 2
1.3 The contributions of this work 3
1.4 Organization of this thesis 4
2 Theoretical Background 5
2.1 Theory of Planned Behavior 5
2.2 The Stress Staircase 7
2.3 The Zone of Proximal Development 9
2.4 Activity Theory 10
2.5 Transition Systems 11
2.6 Action Reasoning 11
3 Related Work 13
3.1 Prior research in human-aware planning 13
3.2 Theory of planned behavior in intelligent systems 14
4 Methods 16
4.1 Knowledge Elicitation 16
4.2 Knowledge Modeling 17
4.3 Knowledge Incorporation 17
4.4 Prototype using the DLV-k language and Unity 3D 17
5 Results 19
5.1 Knowledge Elicitation: Children with autism 19
5.1.1 Signs of stress 19
5.1.2 Sources of stress 21
5.1.3 Management of stress 22
5.3 Representation of the Environment 29
5.4 Goal reasoning with dynamic goals 31
5.5 Intelligent System Architecture 34
5.6 Evaluation of Human-aware Planning 42
6 Discussion 47
7 Conclusions and Future Work 52
References 53
1 Introduction
This study aims to find ways for an intelligent software agent to understand a human’s be- havior in a situation, and in this process, plan its own actions in order to assist the human in reaching their goals. A situation where the human feels anxiety, such as a social situation, can be simulated in a Virtual Reality (VR) environment. The variables of the VR envi- ronment can then be dynamically adapted by the system; In this way assisting the human through difficult parts of the scenario, with the aim of promoting behavior-change.
The use-case is a VR game, developed for children with autism for practicing social scenar- ios, scenarios that children with autism often find stressful or scary. This study investigates how an intelligent software agent can, by continuously analyzing how the user interacts with the simulated situation, adapt the simulation in order to personalize the experience.
Cognitive theories of human-behavior and learning are explored in order to define a com- putational model for promoting behavior-change. In this approach, it must be found; what interventions to be performed by the system; how these interventions should be performed;
and when these interventions should be performed, in order to effectively promote behavior- change. An action reasoning computational model is proposed inspired by three cognitive theories; Theory of Planned Behavior [4], answering the what and how questions of the model; the Stress Staircase [16] and the Zone of Proximal Development [38], together an- swering the when questions of the model. In order for the child to succeed with social challenges in the Virtual Reality scenario, the system does adaptations of the environment for providing assistance. The variables of the environment and the human’s mental state constitutes Interaction Constraints (IC) [28] for the system. This defines a problem con- cerning human-aware planning [8].
1.1 Human-aware planning
Human-aware planning is a way to improve the ability of autonomous systems to plan its ac- tions in a space that is populated and affected by humans [8]. In human-aware planning, the intelligent system is required to make alternative hypotheses of the humans’ plans, i.e., pre- dict the actions that the human might perform in the future [13]. In addition, the intelligent system is required to manage human goal achievement [13]; understand the relationship between human goals and which partial goals that are required to be achieved for reaching a future goal. This study aims to explore this human-aware planning problem in a scenario where the agent’s task is to promote human behavior-change, in order to assist the human in reaching goals. In this scenario, the dynamic sub-goals of the human must be kept in the agent’s planning process while pursuing the human’s future goals.
In any human activity, the human has goals which they aim to achieve. The human’s plan consists of a sequence of human actions with the aim to bring the human closer to these
goals [8]. In a scenario where the human activity, in addition, involves an intelligent agent which has its own specific goals and plans that run in parallel with the human, the scenario quickly rises in complexity regarding how these actors can interact meaningfully. Human- aware planning involves the relationship between the human and the intelligent agent; the interrelationship between their goals and plans in the activity [8]. The human’s plan can be recognized and predicted by the intelligent agent through observations of the environment and the human’s actions, in this way predicting the human’s future plan. In human-aware planning, the artificial agent strives to cooperate in various ways with the human, e.g., to assist the human in an ongoing activity. The agent’s plan, in the same way as for the human, consists of a sequence of actions. These actions must be planned in relation to the predicted human plan and other constraints of the environment. The various actions that the human takes in a situation, as well as preferences, regulations and social norms that are present constitute Interaction Constraints (IC) [28] that limit and control what actions the agent can take in a given situation to best interact with the human. This is done by adapting the agent’s plan to the situation by constantly observing the world and updating the agent’s knowledge.
When the predicted plan of the human changes, the agent has to change its plan accordingly.
The agent has programmed models of the environment, of the human, and of the human’s activities, a knowledge base which is comprised of constraints and relationships between activities, actors and objects in the current environment.
The complex situation where the human plan and the agent’s plan are to be interwoven is a problem regarding goal reasoning [21]. The human has a final goal, but along the way there are a series of sub-goals that may lie in a sequence, successively pursued to finally reach the goal. However, the sub-goals can also be mutually exclusive, more or less necessary, recurring, ongoing, temporary, and with certain probability due to the stochastic human behavior. A problem in human-aware planning is whether these dynamic sub-goals of the human can be preserved in the agent’s predictive models of the human; e.g., in a situation where an intelligent agent aims to assist the human to achieve the sub-goals, and eventually to reach the final goal. This requires a model that can identify the current situation, identify which sub-goals that are present in the human’s plan, identify which sub-goal the human is currently pursuing, which goals are achieved and which are coming next.
1.2 The aim of this thesis
This study aims to introduce computational models to deal with scenarios where the intel- ligent agent’s task is to promote human behavior-change, assisting the human in reaching their goals in a virtual reality environment. In these virtual scenarios, dynamic sub-goals of the human must be kept in the agent’s planning process while pursuing the human’s higher goal. Given this human-aware planning problem, the following research questions arise:
Q1. By understanding the environment as an agent, how can the agent reason about dynamic goals in human-aware planning?
Q2. How can the environment be adapted in order to promote human behavior-change?
In order to approach the human-aware planning problem, and to answer the research ques- tions, the study has the following objectives:
Objective A. Develop an intelligent software architecture that can reason about human
plans while preserving dynamic human goals, and generate an agent plan on how to promote human behavior-change, assisting the human in reaching their goals. This objective entails exploring theory of planned behavior as a theoretical base, and to develop an architecture utilizing an action reasoning theory.
Objective B. Model the environment for a practical use-case by conducting knowledge elicitation with domain experts. Objective B entails to specify the user model to make the system behave in a way appropriate for the human-aware scenario. This is done in an qualitative approach with data collection through interviews. The use-case is a virtual reality game, developed for children with autism for practicing social scenarios. The domain experts, e.g., psychologists, are interviewed to identify key elements in the environment;
actors, objects, events, relations and constraints.
Objective C. Integrate and test the architecture in the practical use-case, where a human and the agent interact to achieve the human’s goals. This objective entails to do an evaluation of the architecture with a set of domain experts, e.g., psychologists, in order to find the usefulness, strengths and limitations of the approach, and to find directions for future work.
1.3 The contributions of this work
An action reasoning computational model is proposed modelled in accordance with the Theory of Planned Behavior (TPB) [4], a cognitive theory to explain and predict an individ- ual’s intention to engage in a behavior at a specific time and place. The general idea is that the individual’s beliefs about a behavior have causal effects on the individual’s attitudes, subjective norms, and perceived behavioral control in the behavior, which in turn promotes or inhibits engagement in the behavior. According to TPB, perceived behavioral control and behavioral intention are direct predictors of behavioral achievement. Due to the causal structure of this theory, and its empirical grounds for being able to predict human behavior, it can be a promising framework for human-aware planning.
In addition, two other relevant cognitive theories where found throughout the study; the Stress Staircase [16], explaining levels of stress in children, and the Zone of Proximal De- velopment [38], explaining how and when learning best is achieved. Theory of Planned Behavior, together with the two other cognitive theories can constitute a promising frame- work for the human-aware planning problem of the current study. These theories are further explained in Section 2, and discussed in relation to a computational model in Section 5.
In order to formalize a computational model, Answer Set Programming (ASP) [31] is uti- lized. ASP is a declarative problem solving paradigm that has its roots in Logic Program- ming and Non-monotonic Reasoning. Through the ASP solving process, a plan to reach the next state or goal is generated which strictly follows the specified constraints of the environment. This study explores how ASP can provide a plan that assists an individual to accomplish their goals in a scenario. To reason about the goals of the human, the agent perceives the current state of the environment and generates an action plan based on the specified constraints; e.g., temporal and sequential constraints between activities, human preferences, social norms, and by perceiving and identifying the human’s mental and phys- ical state. In this way, adapting the environment to make the tasks easier for the individual.
By defining the user’s physical and mental state based on fluents of the environment, plans
can be generated for providing goal-oriented interventions. A final step in the proposed ar- chitecture is the conduction of figurative actions. The system creates a figurative instance of the scenario, an evaluation process in which the system imagines the user’s future behavior.
In this evaluation process, the alternative plans generated by the ASP planner are evaluated in relation to weights tailored for the individual user, selecting a final plan for adapting the virtual scenario.
1.4 Organization of this thesis
The rest of this thesis is organized by first briefly presenting the state-of-the-art in human- aware planning and goal reasoning. The methodology is then presented, which describes (1) the knowledge elicitation process with domain experts, (2) the knowledge modeling through a theoretical framework; Theory of Planned Behavior, and (3) the knowledge incorporation process, i.e., the formulation of an intelligent system architecture incorporating the domain knowledge. The result is then presenting an overview of the knowledge modeling process and a specification of the proposed intelligent system architecture. The thesis is concluded by a discussion of the architecture’s potential, limitations, possible use-cases, and directions for future work.
2 Theoretical Background
This section explains the cognitive theories that have influenced the modeling of the pro- posed intelligent architecture; Theory of Planned Behavior [4], the Stress Staircase [16], and the Zone of Proximal Development [38]. Activity theory [27] is then explained, a cognitive theory regarding human activities, which has structural relations to the proposed architecture for providing a goal-oriented structure. Computational paradigms of the pro- posed architecture are finally explained; Transition systems [20] and Action reasoning [20].
2.1 Theory of Planned Behavior
Theory of Planned Behavior (TPB) [4], later reformed as The Reasoned Action Approach (RAA) [18], is a cognitive theory explaining and predicting an individual’s intention to en- gage in a behavior at a specific time and place. The general idea is that the individual’s beliefs about a behavior have causal effects on the individual’s attitudes, subjective norms, and perceived behavioral control in the behavior, which in turn promotes or inhibits en- gagement in the behavior. The key component in this model is as such behavioral intentions which are influenced by (1) the individual’s attitude about a behavior, (2) the individual’s concern about social norms in relation to the behavior, and (3) the individual’s perceived behavioral control in conducting the behavior.
These components are grounded in the individual’s beliefs about the likelihood that the behavior will have the expected outcome. If the behavior is conducted, what will be the consequence? Is the behavior socially acceptable by others? Does the person evaluate their own skills and self-efficacy to be high enough for conducting the behavior? According to TPB, perceived behavioral control and behavioral intention are direct predictors of behav- ioral achievement (see Figure 1).
Figure 1: The Theory of Planned Behavior. An individual’s beliefs about a behavior lead to attitudes, subjective norm and perceived control towards the behavior. These three evaluations are predictors of the intention towards the behavior. Behav- ioral achievement can then be predicted based on the behavioral intention and the perceived behavioral control. A final component influencing the behavioral achievement is the individual’s actual control. A person is reflecting upon past experience of actual control, which influences the perceived behavioral control in future behavior. This figure is adapted from [3].
Attitude (A) refers to the degree to which an individual has a positive or negative evaluation of the behavior. This entails a consideration of the outcomes of performing the behavior.
The overall attitude towards the behavior is a consideration of each expected outcome b of the behavior, multiplied with the individual’s perceived value e of that outcome. All expectancy-value pairs in the behavior are then summarized, resulting in the overall attitude towards the behavior (see Equation 2.1).
A ∝
∑
bi∗ ei (2.1)Subjective norm (SN) refers to the belief about whether people approve or disapprove of the behavior i. The individual’s beliefs about what people of importance to the person think of their, or the individual’s, engagement in the behavior. This involves social norms, which are normative behavior for a group of people. The overall subjective norms towards the behavior is a consideration of each normative belief n of the behavior, multiplied with the individual’s motivation m to comply with that norm. All normative-motivation pairs in the behavior are then summarized, resulting in the individual’s overall evaluation of subjective norms towards the behavior (see Equation 2.2).
SN ∝
∑
ni∗ mi (2.2)Perceived behavioral control (PBC) refers to an individual’s perception of the ease or diffi- culty of performing the behavior i. Perceived behavioral control can vary for an individual depending on the activity or action that is to be performed. The overall perceived behavioral control towards the behavior is a consideration of each performance p aspect of the behavior, multiplied with the individual’s perceived controllability c of that aspect. All performance- controllability pairs in the behavior are then summarized, resulting in the overall perceived behavioral control towards the behavior (see Equation 2.3).
PBC ∝
∑
pi∗ ci (2.3)According to TPB, behavioral intention (BI) is the motivational factor that influences a given behavior, the likelihood that the behavior will initiate in the behavior. Behavioral intention is the sum of the overall attitude, subjective norm, and perceived behavioral control. Specific activities can be more or less affected by these three predictors. Thus, an activity specific weight w is added to each predictor (see Equation 2.4).
BI= w1A+ w2SN+ w3PBC (2.4)
Finally, behavioral achievement (BA) can be calculated, the likelihood that the individual achieves the behavior. Behavioral achievement is the sum of the behavioral intention (BI) and perceived behavioral control (PBC).
BA= BI + PBC (2.5)
2.2 The Stress Staircase
The Stress Staircase [16] is a model and methodology used by special education teachers to define and recognize stress levels in the children that they work with. The methodology specifies tools or measures for every level in the stairs on how to best approach the child for calming directions, training abortion, and application of different stress managing strategies specific for the individual child.
Stress and anxiety, and behavioral problems as a result, are common in children with autism and related disabilities. This stress is often triggered by stressors in the environment such as having to expose themselves to challenging social situations. With limited ability to shield themselves from these stressors, the child is easily overwhelmed. It is therefore intentional to be able to identify the child’s stress level, to be able to calm the child early, and to set requirements based on their current ability.
Special education teachers and other experts working with autistic children often find that these children exhibit aggressive out-acting behavior without any warning whatsoever. In any case, it has been shown that the children go through a series of states before they reach this state of execution. Where at each stage they aggregate this stress. If the child receives the necessary support early in the process, this process can be stopped before it goes too far. However, it is common for these children to go through the whole process without even realizing that they have been under stress. It is therefore important for those who work with these children to understand this step-by-step process, learn to recognize these steps and find ways to help the child in the step they are on.
The stress staircase is a method for defining these steps. The first steps show early signs of stress, often in the form of restless behavior; The child gradually increases in stress, which often manifests itself as lack of self-control, compulsive behavior, aggression and out-acting where the child is usually unreachable. The stress staircase divides the process into several steps that go steadily upward according to the steps: balanced, unbalanced, compulsive, explosive, out-acting, and ends in a step of mental ill-health. The method is based on
theories regarding how the child should be accommodated depending on what stage the child is on. It is emphasized that the child should be met with calming directions as early as possible in the process. At each step the possibility of diffusion drastically decreases. At a certain point, not much can be done but to let it run its course (see Figure 2).
Figure 2: The stress staircase. The child’s stress is classed according to a level structure.
Starting at a state of balance, going step-by-step up the stairs as the child’s stress increases. At each step, certain way of approaching the child are appropriate. An individualized stress staircase is learned for each child. This figure is inspired by [16].
When the stress is low, the child can be exposed to learning activities, while on higher levels the child must be relieved of demands and expectations. The stress staircase is thus a way of identifying the child’s current ability, how the child should be approached, and what level of challenge or activity the child may be exposed to at the moment. Special education teachers apply this method as a strategy to set requirements based on ability for the child to always succeed.
The stress staircase is interesting from a computational perspective for recognizing stress levels. If ways can be found for recognizing which stress stair a child is on, appropriate approaches to stress relieves can be suggested by a system. Prior studies have proposed interactive environments with sensors for monitoring stress levels. These techniques have generally relied on physiological signals, e.g., electroencephalography (EEG) [26], heart rate variability (HRV) [1], physical signals, e.g., blink rate [23], or behavioral measures, e.g., gestures [30] and body movement patterns [37] with the help of mobile sensors.
2.3 The Zone of Proximal Development
The zone of proximal development (ZPD) [38] is a condition where, e.g., a child has achieved a learning goal and is about to learn something new. At this stage, the teacher or a more knowledgeable person can guide the child. According to the theory, with a little assistance from the surroundings, a child can often solve problems that would otherwise be difficult to solve alone. The ability to learn is however linked to the child’s individual level of development. Thus, learning cannot happen if the degree of difficulty is far from the child’s zone of proximal development. Initially, the teacher gives a lot of support and help, and gradually the assistance decreases or ceases completely as the child can learn the skill them self. In this way, the learning process goes in a cycle between receiving assistance and working independently. The cycle moves upwards in difficulty, in line with the child’s achieved learning goals and acquired skills (see Figure 3).
Figure 3: The zone of proximal development. The child can together with a teacher handle new challenges. The child can then proceed alone on a stable difficulty, until the child is ready for the next level together with a teacher. The zone of proximal de- velopment lies in a level of difficulty that is not too easy, even without assistance, and not too difficult, even with assistance. This figure is inspired by [38].
According to ZPD, learning is a social embodied construct, conducted in interaction with the environment. In this way, the child takes control of their own learning through exploration.
In this environment, multiple zones can exist simultaneously that are interrelated in the environment; different learning goals, or challenges, relating to each other by the constraints of the environment. In this way, different aspects of the environment can be more or less difficult, but in the collective manner supporting learning by letting the child lean on aspects that feels easier, while working on the more difficult aspects.
ZPD can be interesting from a computational perspective for modeling learnability, rec- ognizing appropriate learning levels, i.e., challenge levels. Computational approaches to model ZPD have been made in prior studies for measuring and predicting learning capabil- ities, e.g., in order to maintain the student in their ZPD. In [12], predicted probabilities of correctness are used while students engage in reflective dialogue. In their proposed system, an Additive Factor Model (AFM) [12] is used to model students’ knowledge. The student model attempts to predict if a student is able to give a correct answer without assistance.
In [5], an analytical model is proposed that captures the learning capabilities of a learner.
Their model is formally presented in the declarative logic programming language of DLV.
The proposed model provides a representation of the learner’s structural knowledge fron-
tier and social knowledge frontier, i.e., the learner’s ZPD. Based on the identified ZPD, the agent can propose appropriate learning tasks for maintaining the learner in ZPD.
2.4 Activity Theory
Activity Theory is a conceptual framework forming a hierarchically-structured graph for representing and reasoning about human activities. An activity is comprised of a set of sub-activities, and each sub-activity is in turn comprised of actions and operations at a low level. A central concept is that each sub-activity in this structure is directed to fulfilling a goal, and the overall activity is in turn oriented towards the activity motive, the higher goal.
In this goal-oriented structure, sequential relations and constraints between activities can be represented (see Figure 4).
Figure 4: Activity Theory forms hierarchical representations of human activities. Activi- ties are goal-oriented structures comprised of sub-activities. Each sub-activity is directed towards a goal, and the overall activity is oriented towards a motive. On a low level, the activity is conducted through routine-based operations that are influenced by conditions of the environment. This figure is adapted from [27].
An activity is defined as an interaction between a person and their environment. In the pro- cess of conducting an activity, a person use physical and psychological ”tools” to mediate the world, in direction towards the goal [27]. Actions are conducted through operations, low level actions specific for the individual. Operations are routine-based processes that provides an adjustment of an action to the ongoing situation. The operations are oriented toward the conditions under which the individual is trying to achieve a goal, i.e., condi- tions in the transitions between goals. The individual is often not aware of their operations.
Instead, operations are conducted based on routines, instincts, motivators, stressors, fears, etc. These routine-based operations are discussed more in Section 5.4, in regard to goal reasoning of the proposed architecture for human-aware planning.
2.5 Transition Systems
A transition system [20] consists of a set of configurations and a set of transitions. Transition systems are used to describe dynamic processes with configurations representing states and transitions saying how to go from state to state. These transitions are the result of actions executed in a specific state, leading to a new configuration that represents another state. A transition system has a finite set of states (the state space). In the state space, a set of initial states and a set of goal states are defined. A transition system can, through its predefined action signatures and causal rules, generate plans on how to reach a goal state from an initial state. A plan is a sequence of actions such that each action changes values of fluents, eventually resulting in a goal state.
In a transition system, an action signature consists of a set of value names V, a set fluent names F, and a set of action names A. Any fluent, environment variable, in F is represented by a value from V in any specific state, which is a collection of fluents configured with specific values. An action a, if executed in some state s, leads to a resulting state s’. In this process the action a is changing the values of the fluents in state s, ending up in the new configuration, state s’. A transition system is deterministic if there is only one initial state and all actions are deterministic. Hence all future states of the world are completely predictable. However, the resulting state may not be uniquely determined by the initial state and a specific action. Thus, a transition system can also be non-deterministic.
A transition system can be represented as a labeled directed graph. Each state s is repre- sented by a vertex linking fluent names to value names. Every triple <s, A, s’> is repre- sented by an edge leading from state s to state s’, i.e., the transition from state s to state s’
by executing action A (see Figure 5).
Figure 5: A transition system can be represented as a labeled directed graph, where each node is a state, and each edge is a transition between states, e.g., the transition from state s to state s’ by executing action A that changes the value of fluent f1 from true to false.
2.6 Action Reasoning
Action reasoning regards the logical descriptions of the actions that are defined in a tran- sition system. When can an action be executed and what will be result after the action is executed in different conditions.
Actions are defined by a set of preconditions and post-conditions. Preconditions specify when actions are executable, i.e., in which configuration of fluents can an action be exe- cuted. Post-conditions specify which fluents that will change after the action is executed.
Each action has a set of causal effects defined by causal laws. A causal law of an action is specified in the form:
acauses ( f ) if e0, ..., em.
The law says that action a, executed in a state that satisfy the conditions e0, ..., em, cause the variable f to change.
The preconditions are specified as constraints that tell when the actions can be executed. An executable constraint for an action is specified in the form:
executable (a) if e0, ..., em.
The law says that action a can be executed in a state that satisfy the conditions e0, ..., em, i.e., that specific fluents are present with the specified values.
Both static and dynamic causal laws can be defined. Dynamic laws specify causal effects after an action has been executed, while static laws specify causal effects in relation to conditions of fluents that are not restricted to a specific action. The static laws are always executed if the specified configuration of fluents are met. This makes it possible to define indirect causal effects of actions.
An action language is used for specifying state transition systems, explained in Section 2.5 above, and are commonly used to create formal models of the effects of actions on the world. Action languages are commonly used in the AI and robotics domains, and are, e.g., used for automated planning [20].
3 Related Work
This section presents prior research in the area of human-aware planning, and approaches for making comprehensive frameworks for these systems. This section then presents prior research where the Theory of Planned Behavior [4] has been used for modeling intelligent systems. The section is concluded by motivating the approach of the current study in rela- tion to prior research.
3.1 Prior research in human-aware planning
The human-aware planning problem has in general been explored in scenarios where a robot is situated in an environment involving humans, where the robot must incorporate the mental model of the human into its planner’s deliberative process. Thus, the planner must in addition to the planner’s own model take into account an estimate of the human model.
The general solution for this planning problem is a joint plan where the robot plan adapts to meet the requirements of the human plan [9]. A survey of recent efforts in human-aware planning is presented in [9] and in [2]. A sample of prior work in this direction is presented next.
A prior study [28] explores human-aware planning and addresses the challenge of automat- ically generating plans that have to accommodate scheduled activities, features and pref- erences of humans. The use-case is a scenario where a set of robots have to plan for a whole day in a household environment that is co-inhabited by a human family. To handle such domains the study proposes to use interaction constraints to model how robot plans and human activities should relate to each other. The planning algorithm uses causal rea- soning to create a plan using heuristic forward planning together with a Causal Graph [24].
Temporal propagation [14] on goals is used to determine an order and extract a structure of requirements between goals which enforces that goals that have an earlier start time have to be achieved first. The study suggests a direction for future research to extend interaction constraints to allow new goals as resolvers, where an autonomous system pursues a goal, in response to the human activity, that is not one of the original goals.
A prior work [19] explores goal reasoning agents that are able to dynamically reason about their goals, and modify them in response to unexpected events or opportunities. The work describes a hybrid goal selection and planning approach for goal reasoning agents. The approach allows for agents that are members of human-agent teams to use the partially specified preferences of the human to estimate the utility of goals and guide goal selection.
The work presents a goal reasoning agent that can commit to multiple goals concurrently.
In addition, the agent is able to use the partially specified preferences of the human in its reasoning process. The agent maintains a set of selected goals that represent the goals it is currently attempting to achieve. During agent plan execution, at any time the agent can add a goal, remove a goal, or abandon all previous goals and commit to only a single goal. The
work utilizes SapaReplan framework [43].
A prior work [13] proposes human-aware planning as a way to improve the ability of robots to coexist with humans. The work presents HA-PTLplan, a human-aware planner designed to be part of a larger framework deployed in a real environment, and explores a use-case in which a cleaning robot interacts with a human in a home environment. This planner is able to take into account forecasted human actions at planning time, and has the possibility to generate policies with partial goal achievement. The planner takes a set of possible human agendas, the interaction constraints, and a set of goals as input and generates a policy for the robot that is compliant with the inputs. The work proposes future work that expands the planner by introducing human activities with uncertain temporal duration.
3.2 Theory of planned behavior in intelligent systems
Approaches have been made to model intelligent systems based on the theory of planned behavior (TPB) [4]. Efforts in this direction have generally focused on agent simulations.
However, approaches for Human-AI interaction, e.g., persuasive systems based on TPB have also been explored. A sample of these works is presented next.
A prior work [10] reviews developments in computational persuasion, with a particular fo- cus on domain modeling. An approach for domain modeling for computational persuasion is proposed. The core of this proposal is an ontology based on particular kinds of be- lief, identified as being important in behaviour-change. The proposed approach for domain modeling is intended to facilitate the acquisition and representation of arguments that can be used in persuasion dialogues, and the strategic choices of arguments to present. The work explores behavior-change theories for the domain modeling, one of these theories being the theory of planned behavior. Prediction of behavior-change is made by acquiring data regarding the persuadee’s attitude towards the behaviour, subjective norms, and behavioral control. The proposed concept is discussed in the context of an automated dialogue system.
The system finds out information about the persuadee’s beliefs through multiple-choice questions presented on a mobile phone interface. The system then presents arguments with the aim of changing these beliefs.
A prior work [35] presents a computational implementation of the Theory of Reasoned Action, the predecessor of the Theory of Planned Behavior, using artificial neural networks.
In the proposed model, behavioral intention arises from a dynamic constraint satisfaction mechanism among a set of beliefs. In a set of simulations, the paper shows that constraint satisfaction can simultaneously incorporate the effects of past experience with the effects of immediate social context to produce behavioral intention. The paper presents the predictive ability of the model with respect to empirically derived behavioral intention. The paper represents a significant advance in theory towards understanding the dynamics of health behavior.
A prior work [25] presents automated evaluation of human behaviour. A model for behav- ioral AI is presented designed with the aim of being able to play rationally, adhering to formally stated behaviour preferences, and ensuring that very specific circumstances can be forced to arise within a game. The work is based on established cognitive models, formal logic, and approaches from game theory. One of these models being the Theory of Planned Behavior. The proposed model for behavioral AI has been implemented in a computer game
to support AI players that exhibit specific behavioral preferences.
In contrast to typical applications of human-aware planning, where a robot is situated in an environment populated by humans [28], where it needs to adapt its actions in relation to the environment, this study has explored how a software agent can understand the human’s behavior in an environment, and generate a plan on how to adapt the environment in order to promote human actions. The agent creates a model of the human, and utilizes action reasoning to deliberate about the human’s future actions and beliefs. The human model is based on the theory of planned behavior, capturing contextual motivations for the human’s actions. This study is novel in its approach of human-aware planning for promoting human behavior change.
In the current work, the agent does not plan actions for the human, but instead tries to predict actions and plans of the human based on the state of the environment. The planning approach of this work differs from typical approaches of plan merging [22] between robots.
In the use-case of this work, in contrast to robots, the human is not controllable; the agent can only adapt the environment attempting to promote human actions. Furthermore, the human’s actions can lead to new goals for the agent. A consequence for this is that the plan of the robot cannot be created in advance. Thus, the model must account for dynamic goals in its deliberative process.
4 Methods
This study explores human-aware planning and goal reasoning in a VR game scenario. The VR-game is designed for children with autism, who can through the virtual environment practice social scenarios. The virtual environment is fully observable, meaning that the user’s position and movement in the environment can be acquired, as well as positions of virtual objects and avatars in the user’s surroundings. In addition, through sensors the human’s pulse and eye movements can be acquired. In this way, the agent can identify the current state of the environment and the human.
The study concerns (1) knowledge elicitation, i.e., understanding behaviors, actors, rela- tionships and interaction constraints involved in the specified use-case, (2) knowledge mod- eling, i.e., making the acquired knowledge explicit in relation to the theory of planned behavior, and (3) knowledge incorporation, i.e., formalizing of the intelligent system that incorporates the knowledge.
In this process answering research question Q1 by exploring the architecture of the intelli- gent system and its human-aware planning.
Finally, the human-aware algorithm must, to be complete, be evaluated with actual humans in the loop. Research question Q2 is answered through the findings of a controlled user study with the aim to evaluate the usefulness of this human-aware planning approach. This is done through a qualitative research study with a set of domain experts, e.g., psychologists, participating in interviews and prototype sessions. These sessions will collect the experts’
thoughts of the behavior of the intelligent agent in the VR-scenario. Appropriate feedback from the intelligent system and the model of the environment is explored in the provided use-case. See Section 4.4 for a description of the intelligent system prototype.
4.1 Knowledge Elicitation
The study follows an interpretive approach to find epistemological and ontological expla- nations of the target phenomenon, i.e., the behavior of the target users and their relationship with elements in the social environment. Both inductive and deductive perspectives are made throughout the analysis. The knowledge elicitation is deductive in the way that it re- lates the data to a theoretical framework, theory of planned behavior. This is done in order to give the approach structure through an established cognitive theory. The study is also inductive, where data is analysed in a bottom-up manner, using data from the interviews to build codes, categorizations, and themes to generate relevant interaction constraints of the specific use-case.
Data collection. Qualitative semi-structured interviews with domain experts in order to find explanations of the target users’ behaviors in relation to the social context. These interviews are focused on the experts’ experience in working with autistic children.
Data analysis. A Thematic analysis is conducted in which the interview transcripts are coded, categorized and themed. The analysis is iteratively revised by constant comparison with new codes emerging in further data collection.
Selection of Participants The participants of the user study are considered domain experts working with children with autism. A requirement is to have experience working with these children, and preferably to be currently active in this work. See the complete selection of participants in Table 1.
Table 1 Participants for the knowledge elicitation process.
Participant Occupation Experience Gender Session
P1 Psychologist 30+ Years F Interview
P2 Psychologist 5 Years F Interview
P3 Psychologist 2 Years F Interview
P4 Special-Education Teacher 13 Years F Interview
P5 Special-Education Teacher 4 Years F Interview
4.2 Knowledge Modeling
When constructing the computational model, a theoretical framework for human behavior, the theory of planned behavior, is explored as a theoretical base. In this way, relating the different elements of the environment in a way that more accurately represents constraints and relationships that are relevant in human behavior change. A computational knowledge base is developed, which is presented in Section 5 and further discussed in Section 6.
4.3 Knowledge Incorporation
In order to evaluate the computational knowledge base, a proof of concept prototype is de- veloped. The prototype follows the high level computational model inspired by the cogni- tive theory of behavior change, the theory of planned behavior. The proposed architecture is utilizing action reasoning based answer set programming to represent the constraints of the environment, states, goals, and causal rules for actions to adapt the environment. The intel- ligent system analyzes the human interactions with the environment, consults its knowledge to generate a plan, and responds by adapting the environment. In this way, the intelligent infers tailored actions to the user in reaching their goals. The architecture is presented in Section 5 and further discussed in Section 6.
4.4 Prototype using the DLV-k language and Unity 3D
A proof-of-concept prototype of the proposed architecture has been developed in DLV-k.
DLV-k [17] is a system for planning under incomplete knowledge. The DLV-k language is capable of modeling transitions between states of the world, i.e., states of complete knowl- edge, and reasoning about them as a particular case [17].
A planning problem in DLV-k consists of (1) a planning domain that specifies the rules of the domain, (2) a query that specifies the goal, and (3) a set of background predicates, which is a collection of fluents from the domain.
A planning problem in DLV-k is translated into an equivalent disjunctive logic program.
The solutions of the original planning problem, the generated plans, are finally obtained from the answer sets produced by DLV and returned as a sequence of actions [17].
The virtual environment is created in the Unity 3D platform. The Unity 3D application collects background facts based on the current state of the 3D environment. These facts are formatted to an Answer Set Programming (ASP) input string that is sent to the DLV-k program. The returned sequence of actions are filtered and collected, and finally executed to adapt the virtual environment.
The virtual environment represents a school cafeteria (see Figure 6). The user is entering the lobby, walking around the corner to find the food counter. At the counter the user can pick food items and then greet the cashier before finding a table. While the user is walking through the virtual environment, the intelligent software agent is receiving facts about the user’s state and the current configuration of the environment. The agent plans appropriate adaptations which are finally executed. In the current prototype, stressors in the environment, e.g., sound level and the amount of people are adjusted, and arrows can be provided to guide the user.
(a) Cafeteria Cashier. (b) Cafeteria environment.
Figure 6: The virtual environment represents a school cafeteria built in unity 3D. Fluents of the environment are sent to the DLV-k solver to plan adaptions of the environ- ment.
In the next section, the results of the study is presented. This is presented in the form of a discussion leading from initial interview data, all the way through modeling of the data, ending up in a definition of a computational architecture. The conducted methodology of this study aims to result in a comprehensive presentation of a grounded theory.
5 Results
In this chapter, the results of the knowledge elicitation is presented, followed by the model- ing process that leads toward a computational architecture. The aim is to describe the whole modeling process, where knowledge is initially grounded in interview data with experts, and is then through a series of modeling procedures made explicit.
5.1 Knowledge Elicitation: Children with autism
In this section, the knowledge elicitation process, an interpretive approach explaining the social reality that a child with autism experiences, is covered. A focus in the process is on what this study calls elements of stress. The elements of stress consist of (1) signs of stress, (2) sources of stress, and (3) management of stress. In addition, the study aims to find indications on when and how to approach stress, and what actions to take for adapting the environment, i.e., identifying the right time and place for such actions in order to promote social learning and behavior-change.
In a qualitative approach, semi-structured interviews target domain experts in the area of autism in children. The participant group consists of three licensed psychologists (P1, P2, and P3), and two special education teachers (P4 and P5), all with experience in working with young children with autism. The interview process is seen as both inductive and de- ductive. It grounds analyses in interview transcripts, finding a theory in the data, while also scoping the analysis and data collection in terms of the theory of planned behavior in order to shape a knowledge elicitation process to explain human-behavior and behavior-change.
In a thematic analysis, the interview data was categorized, themed and further analyzed. A sample of interview quotes are presented next, followed by interpretation and discussion of the quotes.
5.1.1 Signs of stress
This section presents results regarding signs of stress in autism. The experts were asked about their experience in stress recognition in these children, with special attention to early stress signs and how the stress signs develops. A sample of quotes are presented, followed by interpretations and reflections upon these quotes.
“Every student with us has a stress staircase. Right now we work close to the students, so we get to know them. When we can see that the student is on the rise in stress; it may be, for example, that the student begins to shake his leg, and then he begins to bite into his own sweater. If he is about to do something like this, we will lower the requirements immediately, so that he does not have to go up even further to a level where you eventually become explosive”, said by P3.
“Avoidance behavior is a sign of stress. If it is something that is avoided it can be due to stress or worry”, said by P2.
“It can be anything from the child being self-destructive to becoming motor-anxious, or looking generally stressed. Wiggling or walking back and forth, running, trying to get out of the room, or screaming”, said by P1.
As mentioned by the experts, a common signal of stress in a child with autism is to refuse an activity. It can be characterized by withdrawal or avoidance behavior. This refusal is often because the child does not know what will happen, or how to behave in a certain activity. If the child cannot influence the situation in a controllable manner, the only solution can be to avoid it and to stay away from the location where the activity will be taking place.
The stress staircase
Signs of stress can vary widely and depend on how stressed the child is. It can be phys- ical and behavioral responses such as hesitation, shaking legs, scrubbing with a hand on their body, ripping or biting on their clothes, walking or wiggling back and forth, etc., or emotional responses such as getting angry, sad or violent. For recognizing stress and to classify the level of stress in a child, different methods have been developed. Special ed- ucation teachers who work with children with autism often use a so-called stress staircase to learn to recognize the children’s stress levels, which in turn helps to find an appropriate way to approach the child. The stress staircase is implemented and learned for each child by adapting it to their individual needs. When being able to detect the current stress level of the child, appropriate activities can be chosen for the child that correspond to their current ability, and appropriate measures can be taken in order to calm the child. The general idea is that calming directions from a teacher or parent should be undertaken early on if stress is noticed in the child, as the stress is easier to manage before it rises too high.
“We have a stress staircase for every child. In this way we work with requirements according to ability, so that they always succeed. They should never fail and get angry. If they do not have the ability right now, we will lower the requirements. Something they could do yesterday, they may not be able to do today, because there is something that is not right.
You don’t push them today to do what they could do yesterday. This is the stress staircase
”, said by P3.
“I see immediately when the child is in stress, so knowing the child’s stress staircase is very important. Because I can never make demands when they are in stress, then nothing will work. You know that about yourself when you are in affect. That’s not when you can start discussing things or follow demands”, said by P3.
“The child must succeed, so we provide step-by-step challenges. Challenges can be made into something exciting and fun, done in a comfortable environment with people you trust.
The important thing is that the child succeeds and that we break if it becomes too difficult.
Keep it at a low level, and if it gets too stressful, you immediately back off”, said by P5.
The stress staircase starts at a state of balance and goes in steps up to high stress levels.
These steps usually are: 1) in balance, 2) unbalanced, 3) stressed, 4) compulsive behavior, 5) explosive behavior, 6) out-acting behavior that can be violent or extremely passive, and finally ending in a state of mental ill-health. The experts mention that it is important for the child to always succeed with a challenge. For these children, their confidence is very
fragile, and a failure in an activity can thus do great damage that is difficult to repair. Due to this, the children learn by succeeding, and should never learn by mistakes. The stress staircase helps the teachers with this by providing a way to identify the stress level of the child as well as proper ways of countering it. See Table 2 for a summary of signs of stress.
5.1.2 Sources of stress
This section presents results regarding sources of stress in autism. The experts were asked what kind of stressors, internal or external, that can lead to stress, and if this differs from stress sources in general. A sample of quotes are presented, followed by interpretations and reflections upon these quotes.
“It can be stressful for all children, but it is usually extra stressful for children with autism.
Because they don’t really know how to do things - How to make social contact? What will the response be? And how should I handle that response? It often results in not making contact, because you do not really know how to do things”, said by P2.
“When the demands are too high, when they do not know what is going to happen or what to do. When they finish a task, transitions between different activities is a space of uncer- tainty”, said by P1.
”A requirement or expectation to be social, or that they are hungry or tired, just like the rest of us. It can also be small things, like how someone’s voice sounds or that a keyboard clicks too loudly. It can be just about anything, so you have to examine it for each person”, said by P1.
A common interpretation is that people with autism are stress-sensitive, or that they have less stress resistance. However, a more appropriate description is that they have more stres- sors in their lives that they constantly have to deal with. It is thus not necessarily stress resilience that fails, but the ability to shield oneself from stressful elements, which in turn leads to an overload of stress that the child cannot easily get rid of. Prior studies have shown that people with autism, compared to others, have both higher levels of stress in everyday life and higher levels of stress in the body [33]. These high stress levels further reduce func- tional ability. In many cases, the child cannot cope with the discomfort, which often leads to withdrawal or avoiding behavior. In this way, the child becomes more disabled. Experiences where demands are linked to failures and negative reactions from the surroundings lead to negative self-image and further increase this vulnerability. A major source of stress is thus various demands placed on the child, in combination with the child’s lack of self-image, self-assurance or perceived behavioral control that does not live up to these expectations.
“It is common for the stress to come later. It may be that one student sees another student in affect. This can cause stress for the other student, and then the parents must know what had happened. It is important to know what we encounter, what mood. We try to analyze that all the time. Where they are on the stress stairs”, said by P3.
The problem of slow recovery leads to an aggregation of stress that builds up during the day. This often leads to problems with greater stress levels towards the end of the day, or in an upcoming activity. What has happened earlier in the day therefore has a major impact on the level of stress, as well as the extent to which new stressors affect the child.
Sources of stress can, as stated by the participants, often apply to demands placed on the child. These requirements are often problematic because the child already has a high level
of stress, but also because these children often have difficulties understanding their envi- ronment, especially in new situations. Unpredictability in a situation due to this can lead to stress, e.g., not knowing what will happen, or not knowing what to do. Unknown places, people and situations can thus lead to stress. Social interactions are thus often a major problem. Children with autism generally prefer a structured existence.
Hypersensitivity to sensory impressions can also be more pronounced. Sounds, shouts, lights, and unexpected events, e.g., dropping a plate that is shattered on the floor, can lead to stress. This hypersensitivity can also concern high affective behavior of others. Autistic children are very easily affected by the feelings of others in the environment, e.g., seeing other children in affect. Indirect stress reactions can come in the current situation or later in the day. Due to this, it can be difficult to find the source of the stress, as the source may have come earlier in the day, and is later on confronted. This makes the aggregation of stress an even greater issue. See Table 2 for a summary of sources of stress.
5.1.3 Management of stress
This section presents results regarding management of stress in autism. The experts were asked about their experience in stress management for these children, with attention to how stress can be managed in an ongoing situation. This regards identifying how the child can manage stress alone, what a parent or a teacher can do to help in such a situation, and what stress relieving elements that are in the environment. A sample of quotes are presented, followed by interpretations and reflections of these quotes.
“We would like the child to find environments, or things that cause them to fall into turns.
We know that you feel good about a walk, you feel good about music, that you find the situations that make the student calm. Then the image support is so amazing that they can show what they want to do. Do you want to go for a walk to unwind? Do you want to go to a sleeping room? Go to a listening chair?”, said by P3.
“It is common for the children feel good about walking. Also, riding a bus or car generally makes the children feel good. Partly for controlling the little things. The car becomes a small controllable context. Many young children feel good in a car. It is the sound, the engine that sounds”, said by P3.
“Music can also be such a fantastic thing. We have a student now who has always felt good about having headphones with music. We have now also started to have music during lessons. If he can listen to music, we can put so much higher demands on him. If he listens to music he manages all the steps even in demanding situations. It is absolutely incredible”, said by P3.
“Today she might not go to a cafe. There are requirements according to ability all the time.
It is so special with these young people in the autism spectrum that their world does not look like our world. The goal should be for them to succeed every day”, said by P3.
Stress management is one of the biggest difficulties in autistic children. These children get tired easily due to the high amount of stressors they constantly have to deal with, and thus have a strong dependence on recovery and rest. As the child easily becomes exhausted, it is important to recognize the child’s current ability and present tasks or activities that corre- spond to that ability. When the most important tasks that the child is faced with have been cleared, there should preferably be energy left over for joyful and self-esteem strengthening
activities. As mentioned, when stress is noticed early, it is easier, and a priority, to calm the child as early as possible. Thus, it is important to have a low-affective approach, and introducing requirements based on the child’s current ability in correlation with the stress staircase. This is something that special education teachers always strive for.
Relieve and management of stress is in many ways the opposite of what is causing the stress, e.g., providing structure, for instance through picture support, to handle an unstructured future scenario. Different strategies can be conducted in order to manage stress. These strategies could be learned in advance together with a teacher, or be compulsory stress reactions which in turn help the child to deal with the stress. In this way, signs of stress, sources of stress, and management of stress can interconnect in complex ways.
Predictability in a situation is a way to make a situation less stressful. This can be in the form of preparation with images, or through clear instructions or guides on what to do in a situation. These instructions are often brought to the actual activity, e.g., drawn on a piece of paper. Familiarity and people that the child knows and trusts is a comfort in these situations. The special education teachers often model activities by showing the child how they are done, before the child tries them out themselves, leading to increased predictability.
In a social situation, it is important to find a place that the child feels is safe, together with people that the child is comfortable with, i.e., places and people in the environment that are predictable. It is very common that a child develops such places in areas that are regularly visited, e.g., a school cafeteria. The child can have a specific table, with a specific chair, that feels safe, and always sit next to the same people. This kind of structured environment provides predictability, and when this configuration changes in any way it can lead to stress.
When a stressful situation arises, different strategies can be conducted in order to manage the stress. Some types of distractions are helpful, leading the attention away from the stress- ful activity, e.g., music, that has a soothing effect on most children. Other distractions are to change to the activity, e.g., by conducting a fun physical activity, or by leading the child to a calm environment with less people and less sounds. The special education teachers always strive for a low affective approach towards the children, an important approach for managing a child’s stress due to their sensitivity for high affective behavior of others.
When a child is in stress, it is important to relieve the child from any demands. As men- tioned by both psychologists and special education teachers, it is not a time for requirements or demands when a child is in stress. Introducing challenges with the aim of learning is only suitable when a child is in a state of balance. See Table 2 for a summary of management of stress.
Table 2 Elements of stress. This table summarizes the elements of stress, i.e., signs of stress, sources of stress, and management of stress, identified in the data collection with domain experts. Some signs of stress can also be seen in management of stress, e.g., compulsive strategies. Also, some sources of stress has a corresponding management of stress. In this way, elements of stress can often interconnect in complex ways.
Signs of stress Sources of stress Management of stress
Unresponsive Unknown people Known people
Coping strategies Sounds, Shouts, Lights Favorite music
Shaking legs Social interactions Physical activity
Ripping/biting on clothes Unpredictability Predictability
Scrubbing with hand on body Unstructured Distractions
Walking/wiggling back & forth Not knowing what to do Instructions/Guides
Hesitation Unexpected events Calm environment
Saying “no” High affective (others) Low affective (others)
Withdrawal, Avoiding Demands No demands
Sad, Angry, Emotional Rules not followed (others) Safe place, Safe people
Hitting, Screaming, Violent Unplanned changes Be alone
Self-injurious behaviour Aggregation of stress Coping strategies Increased eye movements Activity transitions Pleasant lights (dim) Increased heart rate
Zone of proximal development
“Facilitate and find the more proximal level of difficulty is important so that the child can successfully learn and practice. It can be too difficult, or too easy, in which case the diffi- culty must be adjusted. Try to settle on a suitable level”, said by P1.
“I am thinking of proximal development zones. You look at the stress levels - If there is no stress at all, then it is not a challenging situation, and then you probably do not practice.
You should have some stress or excitement, for you to practice”, said by P2.
The experts are talking about a theory of learning, the Zone of Proximal Development (ZPD) [38], a concept about when a person is able to learn, stating that people learn by being guided by those who are more knowledgeable and skilled, e.g., with the help of an assistant, and can later continue the improvement by themselves. ZPD is about finding the right level of challenge, which corresponds to the ability of the child. The stress staircase, explained earlier in this section, and the ZPD have important relations regarding when learning should be provided, i.e., when the child is in a state of balance, according to the stress staircase.
According to ZPD, there are a few essential factors that are critical to the success of learn- ing; (1) the help of someone with more knowledge and skills to guide the child, (2) social interactions with the more knowledgeable person that allow the learner to observe and prac- tice their skills, and (3) supportive instructions and tools that help guide the learner through the activity. When the child is in the zone of proximal development, providing appropri- ate assistance and tools gives the child what is needed to accomplish the new task or skill.
Eventually, the instructions and tools can be removed and the child will be able to complete the task independently (see Figure 3).
“Proximal development is how you can be challenged alone or with assistance. But the
