This paper introduces an action language for modelling the causality between a human's motivational beliefs and behavior in human activities. The action language is based on a psychological theory, the theory of planned behavior (TPB), which centers on three sets of beliefs that shape an individual's behavioral intentions: attitude, subjective norms, and perceived behavioral control. The language is modelled in the structure of a transition system in which states correspond to different mental states of an individual, and actions correspond to ways to transition between mental states in order to influence human intentions. We introduce its syntax and semantics. The semantics is characterized in terms of answer sets.
Human-aware planning is a way to improve the ability of autonomous systems to plan its actions
in an environment populated and affected by humans [
Explaining human behavior is a difficult task due to its causal complexity [
A way for reasoning about human behavior is to look at it from the perspective of states and
transitions between states. Action languages, such as A [
By following the concepts introduced in TPB, this work aims to formalize the causality between human beliefs and behavior in an action language. Such a formalization can be utilized by a rational agent to reason about the behavioral intentions of a human agent, and in addition reason about the rational agent’s actions for changing a human agent’s beliefs to influence motivations and intentions.
Modelling the causality between human beliefs and behavior involves identifying which beliefs
and actions, out of many, that cause a behavior. This is a chain of direct and indirect effects that
branch out to change mental states. This is a difficult task that involves the problem of ramification
[
We approach these questions by introducing the action language CT PB. It is modelled in
the structure of a transition system in which states correspond to different mental states of an
individual and an environment, and actions correspond to ways to transition between states in
order to influence human intentions. In order to characterize the semantics of the language
we establish a link between CT PB and Answer Set Programming (ASP) [
To capture human behavior, the action language CT PB has been developed in a knowledge
elicitation process related to a case study [
The rest of this paper is organized as follows. First, we briefly present the theoretical framework; the theory of planned behavior. Syntax and semantics of the proposed action language is then presented which introduces an action reasoning approach to human-aware planning that utilizes the theory of planned behavior. The state-of-the-art in human-aware planning and agents modelling other agents is then presented and discussed. The paper is concluded by a discussion of the action language’s potential, limitations, possible use-cases, and directions for future work.
This section explains the cognitive theory that has influenced the modelling of the proposed action
reasoning framework for human-aware planning; Theory of Planned Behavior [
Theory of Planned Behavior (TPB) [
Attitude (A) refers to the degree to which an individual has a positive or negative evaluation of performing the activity. This entails a consideration of the outcomes of the activity. The overall attitude towards the activity is a consideration of each expected outcome of the activity quantified by the individual’s valuation of that outcome.
Subjective norm (SN) refers to the individual’s belief about whether people approve or disapprove of the activity. The overall subjective norms towards the activity is a consideration of each normative belief of the activity quantified by the individual’s motivation to comply with that norm.
Perceived behavioral control (PBC) refers to an individual’s perception of the ease or difficulty of performing the activity. The overall perceived behavioral control towards the activity is a consideration of each performance aspect of the activity quantified by the individual’s perceived controllability of that performance aspect.
According to TPB, behavioral intention (BI) is the motivational factor that influences a given behavior, the likelihood that the human will initiate in an activity. Behavioral intention is the aggregation 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 empirically derived weight is added to each predictor.
The following subsection presents a qualitative data collection process centered on the components of motivation introduced in TPB. Based on these components, mental states are specified which are captured in a particular transition system.
This section presents the results of a knowledge elicitation process of a related use-case [
By assuming that the current mental state of an individual can be sufficiently recognized, i.e.,
the attitude (A), subjective norm (SN), and perceived behavioral control (PBC), and valuated in
the scale Negative (N) (inhibiting behavior), Medium (M) (indifferent to behavior) and Positive
(P) (promoting behavior), the participating experts were asked how a person’s motivation to a
behavior should be boosted depending on the person’s current mental state. E.g., if the person
currently has a mental state where A is negative, SN is negative and PBC is negative, then which
aspect should be prioritised to boost? If so, what about a similar case where A is medium? By
following this approach, 27 states are specified, each state consisting of the variables A, SN, or
PBC, in which each variable has a value of negative, medium, or positive. Following the experts’
suggestions, we can specify a state space with transition relations between states that represents
a prioritised change of A, SN or PBC depending on the recognized mental state of the human.
A mental state is denoted as A:SN:PBC, labeled by the value of each variable (e.g., PPP if A is
Positive, SN is Positive and PBC is Positive). We can model these relations in terms of a weighted
graph, called a motivation decision-graph (see Figure 1 [
By following the experts’ suggestions of motivational focus in terms of the 27 mental states, a set of trends in the behavior of the system were found: 1. Prioritize to push attitude away from a negative state whenever possible, i.e., to a medium or positive state. 2. Prioritize to push subjective norm away from a negative state, if attitude is at least medium and control is positive. 3. Prioritize to push control away from a negative state if an internal or external motivator is already present, i.e., if attitude or subjective norm is at least medium. 4. If all three aspects are medium, prioritize to push attitude to a positive state. 5. The aim is to reach the state PPP, where all aspects are positive. If this state is reached, then keep the variables steady or make small challenges in attitude, subjective norm, or control by pushing the state to any of the lower states MPP, PMP, or PPM.
These behaviors of the motivation decision-graph is formalized by the semantics of the action language CT PB. The language’s syntax and semantics is presented in the following section. 3. Human Behavior in the Action Language CT PB This section presents the syntax and semantics of the proposed action language, CT PB, following the results of a knowledge elicitation process presented in Section 2.2, in the light of the theory of planned behavior. Action descriptions in answer set semantics is then presented together with an implementation of the motivation decision-graph illustrated in Figure 1, in order to characterize the semantics of the CT PB action language.
The alphabet of CT PB consists of two nonempty disjoint sets of symbols F and A. They are called the set of fluents F and the set of actions A. A fluent expresses a property of an object in a world, and forms part of the description of states of this world. A fluent literal is a fluent or a fluent preceded by ¬. A state σ is a collection of uflents (informal approximation, see Definition 3). We say a fluent f holds in a state σ if f ∈ σ . We say a fluent literal ¬ f holds in σ if f ∈/ σ . Definition 1 (Human-aware alphabet). Let A be a non-empty set of actions and F be a nonempty set of fluents.
• F = FE ∪ FH such that FE is a non-empty set of fluent literals describing observable items in an environment and FH is a non-empty set of fluent literals describing the mental-states of humans. FE and FH are pairwise disjoint. • FH = FA ∪ FN ∪ FC such that FA, FN and FC are non-empty pairwise disjoint sets of fluent literals describing a human agent’s attitude, subjective norm and perceived behavioral control, respectively. • A = AE ∪ AH such that AE is a non-empty set of actions that can be performed by a software agent and AH is non-empty set of actions that can be performed by a human agent.
AE and AH are pairwise disjoint.
CT PB is defined by three sub-languages: an action description language, an action observation language and an action query language.
Definition 2. A human-aware domain description language Dh(A, F) in CT PB consists of static and dynamic causal laws of the following form:
The semantics of a domain description Dh(A, F) is defined in terms of transition systems. An interpretation I over F is a complete and consistent set of fluents.
Definition 3. A state s ∈ S of the domain description Dh(A, F) is an interpretation over F such that 1. for every static causal law ( f1, . . . , fn if g1, . . . gn) ∈ Dh(A, F), we have { f1, . . . , fn} ⊆ s whenever {g1, . . . gn} ⊆ s. 2. for every static causal law ( f1, . . . , fn influences attitude f ) ∈ Dh(A, F), we have { f } ⊂ s whenever { f1, . . . , fn} ⊆ s, and f ∈ FA ∧ ( f (_, Positive, _) ∈ s ∨ f (_, Medium, _) ∈ s), and (∃ fi ∈ FA(1 ≤ i ≤ n) ∧ fi(_, Negative, _) ∈ s, or ∃ fi ∈ FA(1 ≤ i ≤ n) ∧ ∃ f j ∈ FN (1 ≤ j ≤ n) ∧ ∃ fk ∈ FC(1 ≤ k ≤ n) ∧ fi(Medium) ∈ s ∧ f j(Medium) ∈ s ∧ fk(Medium) ∈ s). 3. for every static causal law ( f1, . . . , fn influences subjective norm f ) ∈ Dh(A, F), we have { f } ⊂ s whenever { f1, . . . , fn} ⊆ s, and f ∈ FN ∧ ( f (_, Positive, _) ∈ s ∨ f (_, Medium, _) ∈ s), and ∃ fi ∈ FA(1 ≤ i ≤ n) ∧ ( fi(_, Medium, _) ∈ s ∨ fi(_, Positive, _) ∈ s), and ∃ fk ∈ FC(1 ≤ k ≤ n) ∧ fk(_, Positive, _) ∈ s. 4. for every static causal law ( f1, . . . , fn influences control f ) ∈ Dh(A, F), we have { f } ⊂ s whenever { f1, . . . , fn} ⊆ s, and f ∈ FC ∧ ( f (_, Positive, _) ∈ s ∨ f (_, Medium, _) ∈ s), and ∃ fi ∈ F A(1 ≤ i ≤ n) ∧ ( fi(_, Medium, _) ∈ s ∨ fi(_, Positive, _) ∈ s), and ∃ fk ∈ FC(1 ≤ k ≤ n) ∧ ( fk(_, Medium, _) ∈ s ∨ fk(_, Positive, _) ∈ s).
Definition 4.
1. An inhibition rule ( f1, . . . , fn inhibits a) is active in s, if s |= f1, . . . , fn , otherwise the inhibition rule is passive. The set AI (s) is the set of actions for which there exists at least one active inhibition rule in s. 2. A triggering rule ( f1, . . . , fn triggers a) is active in s, if s |= f1, . . . , fn and all inhibition rules of action a are passive in s, otherwise the triggering rule is passive in s. The set AT (s) is the set of actions for which there exists at least one active triggering rule in s. The set AT (s) is the set of actions for which there exists at least one triggering rule and all triggering rules are passive in s. 3. An allowance rule ( f1, . . . , fn allows a) is active in s, if s |= f1, . . . , fn and all inhibition rules of action a are passive in s, otherwise the allowance rule is passive in s. The set AA(s) is the set of actions for which there exists at least one active allowance rule in s. The set AA(s) is the set of actions for which there exists at least one allowance rule and all allowance rules are passive in s. 4. A promoting rule ( f1, . . . , fn promotes a) is active in s, if a ∈ AH and s |= f1, . . . , fn and all inhibition rules and demoting rules of action a are passive in s, otherwise the promoting rule is passive in s. The set AP(s) is the set of actions for which there exists at least one active promoting rule in s. The set AP(s) is the set of actions for which there exists at least one promoting rule and all promoting rules are passive in s. 5. A demoting rule ( f1, . . . , fn demotes a) is active in s, if a ∈ AH and s |= f1, . . . , fn and all inhibition rules and promoting rules of action a are passive in s, otherwise the demoting rule is passive in s. The set AD(s) is the set of actions for which there exists at least one active demoting rule in s. The set AD(s) is the set of actions for which there exists at least one demoting rule and all demoting rules are passive in s. 6. A dynamic causal law (a causes f1, . . . , fn if g1, . . . , gn ) is applicable in s, if s |= g1, . . . , gn. 7. A static causal law ( f1, . . . , fn if g1, . . . , gn ) is applicable in s, if s |= g1, . . . , gn . 8. A dynamic causal law (a influences attitude f if f1, . . . , fn ) is applicable in s, if s |= f1, . . . , fn , and f ∈ F A, and ∃ fi ∈ F A(1 ≤ i ≤ n), and ∃ f j ∈ F N (1 ≤ j ≤ n), and ∃ fk ∈ FC(1 ≤ k ≤ n). 9. A dynamic causal law (a influences subjective norm f if f1, . . . , fn ) is applicable in s, if s |= f1, . . . , fn , and f ∈ F N , and ∃ fi ∈ F A(1 ≤ i ≤ n), and ∃ f j ∈ F N (1 ≤ j ≤ n), and ∃ fk ∈ FC(1 ≤ k ≤ n). 10. A dynamic causal law (a influences control f if f1, . . . , fn ) is applicable in s, if s |= f1, . . . , fn , and f ∈ FC, and ∃ fi ∈ F A(1 ≤ i ≤ n), and ∃ f j ∈ F N (1 ≤ j ≤ n), and ∃ fk ∈ FC(1 ≤ k ≤ n). Definition 5 (Trajectory). Let Dh(A, F) be a domain description. A trajectory ⟨s0, A1, s1, A2, . . . , An, Sn⟩ of Dh(A, F) is a sequence of sets of actions Ai ⊆ A and states si of Dh(A, F) satisfying the following conditions for 0 ≤ i < n: 1. (si, A, si+1) ∈ S × 2A\{} × S 2. AT (si) ⊆ Ai+1 3. AP(si) ⊆ Ai+1 4. AD(si) ⊆ Ai+1 5. AT (si) ∩ Ai+1 = 0/ 6. AA(si) ∩ Ai+1 = 0/ 7. AI(si) ∩ Ai+1 = 0/ 8. AP(si) ∩ Ai+1 = 0/ 9. AD(si) ∩ Ai+1 = 0/ 10. |Ai ∩ B| ≤ 1 f or all (noconcurrency B) ∈ Dh(A, F).
Definition 6. The action observation language of CT PB consists of expressions of the following form:
( f at ti) (a occurs_at ti) (16) where f ∈ F, a is an action and ti is a point of time.
Definition 7 (Action Theory). Let Dh be a domain description and O be a set of observations. The pair (Dh, O) is called an action theory.
Definition 8 (Trajectory Model). Let (Dh, O) be an action theory. A trajectory ⟨s0, A1, s1, A2, . . . , An, sn⟩ of Dh is a trajectory model of (Dh, O), if it satisfies all observations of O in the following way: 1. if ( f at t) ∈ O, then f ∈ st 2. if (a occurs_at t) ∈ O, then a ∈ At+1.
Let us observe that given a trajectory ⟨s0, A1, s1, A2, . . . , An, sn⟩ where Ai ⊆ A (0 ≤ i ≤ n) and A is a set of actions that can be performed either by a human-agent or a software planner-agent. Actions performed by a human-agent can be movement between areas in the environment, while actions by a software planner-agent are adaptations of the environment that indirectly influences a human-agent’s mental-state fluents, i.e., attitude, subjective norm and control. Definition 9 (Action Query Language). The action query language of CT PB regards assertions about executing sequences of actions with expressions that constitute trajectories. A query is of the following form:
( f1, . . . , fn after Ai occurs_at ti, . . . , Am occurs_at tm) where f1, . . . , fn are fluent literals ∈ FE ∪ FH , Ai, . . . , Am are sub-sets of AE ∪ AH , and ti, . . . , tm are points in time.
This section presents translations in answer set programs of the expressions as part of CT PB.
These expressions incorporate the result from the knowledge elicitation process, specified to
capture explanations of human behavior and behavior-change. Descriptions of expression 1, 9,
10, 11, 14, and 15 follows the definitions in [
In the following translations, the value of a mental state variable, i.e., attitude, subjective norm and control, is represented by h ∈ {Positive, Medium, Negative} in a time point T . Actions are coupled with a planner agent (agent) or a human agent (human).
The above expressions can be modelled according to a specific domain, to capture beliefs of an environment and their influence on human behavior. Accompanying each expression, a domain independent set of rules are applied to filter out solution candidates in a generated trajectory. This is expressed by the predicate motivate which adds a set of integrity constraints based on the motivation decision-graph (presented in Subsection 2.2), defined below.
Definition 10. A motivation decision-graph MDG is a transition system that is a tuple of the form MDG = (M, Act, T, O, MP, L) where M is a non-empty set of states, denoting mental states of a human agent, Act is a set of actions, T ⊆ M × M is a non-empty set of transition relations denoting legal transitions between mental states, O is a set of initial states, MP a set of atomic propositions, and L is a function that defines which propositions ∈ MP valid in each state ∈ M.
In the following definition, we introduce a logic program that characterize the motivation decision-graph of Figure 1. This logic program will help to define the semantics of CT PB. Definition 11. Let PMDG be the following logic program: motivate(attitude,2,1), motivate(attitude,3,2), motivate(control,2,3), motivate(norm,2,4), motivate(norm,3,5), motivate(control,3,6)
The first parameter in motivate/3 corresponds to a mental state fluent, e,g., Attitude. The second parameter in motivate/3 corresponds to the value of a mental state fluent, where value(1) equals to Negative, value(2) equals to Medium, and value(3) equals to Positive. The third parameter in motivate/3 corresponds to a point in time. The motivation decision-graph, prototyped in the above logic program (shared and openly accessible online1), works as domain-independent heuristics for any domain-specific logic program following the CT PB action language.
In order to define the semantics of CT PB, we characterize trajectory models in terms of answer sets. This is formalized by the following theorem: Theorem 1. Let (Dh, Oinitial) be an action theory such that Oinitial are the fluent observations of the dynamic environment in the current state, and the fluents of the currently estimated mental state of the human (Attitude, Subjective norm, and Control). Let Q be a query, according to Definition 9 and let
AQ = {(a occurs_at ti) | a ∈ Ai, 1 ≤ i ≤ m}. 1Source of the motivation decision-graph prototype: https://git.io/Ju9vh
Let T denote the translation of CT PB into a logic program according to the mapping introduced by Section 3.2.
Then, the following statements hold true:
1. If there is a trajectory model ⟨s0, A1, s1, A2, . . . , An, sm⟩ where Ai ⊆ A (0 ≤ i ≤ m) of CT PB
(Dh, Oinitial ∪ AQ), then there is an answer set A of logic program T (CT PB (Dh, Oinitial ∪ AQ) ∪
PMDG, such that for all fluents f ∈ FE ∪ FH at the time points 0 ≤ k ≤ m
(a) holds( f , k) ∈ A , if sk |= f ,
(b) holds(neg( f ), k) ∈ A , if sk |= ¬ f .
(c) holds(occurs(a), k) ∈ A , if a ∈ Ak+1
(d) holds(neg(occurs(a)), k) ∈ A , if a ∈/ Ak+1
2. If there is an answer set A of a program T (CT PB (Dh, Oinitial ∪ AQ) ∪ PMDG and at time
point 0 ≤ k ≤ m
(a) sk = { f | holds( f , k) ∈ A } ∪ {¬ f | holds(neg( f ), k) ∈ A }
(b) Ak+1 = {a | holds(occurs(a), k) ∈ A }
then there is a trajectory model ⟨s0, A1, s1, A2, . . . , Am, sm⟩ of CT PB(Dh, Oinitial ∪ AQ).
Proof 1. (Sketch) Let us start by observing that CTAID [
Let TTAID be the translation of PTAID into a logic program according to [
This subsection exemplifies a use-case of CT PB. Children with autism commonly experience
difficulties in social situations. Lights, sounds, people, and lack of guidance can result in stress and
avoiding behavior. Let us consider an interactive virtual reality (VR) aid for children with autism
for practicing social situations [
Previously, we have done work on context-reasoning (through Activity Theory [
There is a diverse body of research related to the ideas presented in the current work [
Usually, when we talk about theory of mind, i.e., agents modelling other agents [
We have introduced the action language CT PB based on a psychological theory, the theory of planned behavior, explaining a human’s intention to engage in a human activity, and presented how the language can be applied to represent and reason about actions for altering mental-states to promote behavior. By utilizing the action language, an agent can reason about specific human beliefs, i.e., the human’s attitude, subjective norm and perceived behavioral control in a situation. In this way, the agent acquires a particular theory of the mind of the human to deliberate about a human agent’s intentions. On a low level, the language captures a human agent’s beliefs about the environment and how these beliefs correspond to mental states. On a high level, the language captures a suitable direction of motivation based on a human’s current mental state. In this way, a priority of motivation can be utilized for picking the most suitable actions to alterate the environment in order to change the human’s motivational beliefs and influence human behavior. Future work concerns incorporating expressions dealing with probability distributions in human behavior into the language. Furthermore, we aim to explore ways to incorporate customized decision-graphs into the semantics of CT PB, thus going beyond hardwired transition rules. For instance, we aim to develop an emotion decision-graph which extends the semantics of CT PB with models to reason about human emotions and emotional-change.