Figure 2 - uploaded by The Anh Han
Content may be subject to copyright.
Source publication
In this chapter, the authors present an intention-based decision-making system. They exhibit a coherent combination of two Logic Programming-based implemented systems, Evolution Prospection and Intention Recognition. The Evolution Prospection system has proven to be a powerful system for decision-making, designing, and implementing several kinds of...
Contexts in source publication
Context 1
... (Fox-Crow): The BN for Fox’s inten- tion recognition (Figure 2) can be coded with the P-log program in Box ...
Context 2
... sorts bool and fox_intentions , in order to represent Boolean values and the current set of Fox’s conceivable intentions, are declared in part 1. Part 2 is the declaration of four attributes hun- gry_fox , friendly_fox , praised and i , which state the first three attributes have no domain param- eter and get Boolean values, and the last one maps each Fox’s intention to a Boolean value. The random selection rules in part 3 declare that these four attributes are randomly distributed in their ranges. The distributions of the top nodes ( hun- gry_fox , friendly_fox ) and the CPD corresponding to the BN in Figure 2 are given in part 4 and parts 5-8, respectively, using the probabilistic informa- tion pa -rules. For example, in part 4 the first rule says that fox is hungry with probability 1/2 and the second rule says he is friendly with probabil- ity 1/100 . The first rule in part 5 states that if Fox is friendly and hungry, the probability of him having intention Food is 8/10 . Note that the probability of an atom a(t,y) will be directly assigned if the corresponding pa/3 atom is in the head of some pa -rule with a true body. To define probabilities of the remaining atoms we assume that by default, all values of a given attribute which are not assigned a probability are equally likely. For example, the first rule in part 4 implies that fox is not hungry with probability 1/2 . And, actually, we can remove that rule without changing the probabilistic information since, in that case, the probability of fox being hungry and of not being hungry are both defined by default, thus, equal to 1/2 . The probabilities of Fox having intention Food , Territory and Please given the observation that Fox praised Crow can be found in P-log with the queries in Box 2, respectively. From the result of Box 2, we can say that Fox is most likely to have the intention of deceiving the Crow for food, i(food) . There are several ways an EP agent can benefit from the ability to recognize intentions of other agents, both in friendly and hostile settings. Know- ing the intention of an agent is a means to predict what he will do next or might have done before. The recognizing agent can then plan in advance to take the best advantage of the prediction, or act to take remedial action. Technically, in the EP system, this new kind of knowledge may im- pinge on the body of any EP constructs, such as active goals, expectation and counter-expectation rules, preference rules, integrity constraints, etc., providing a new kind of trigger. In order to account for intentions of other agents in decision making with EP, we provide a built- in predicate, has_intention(Ag,I) , stating that an agent Ag has the intention I . The truth-value of this predicate is evaluated by the intention recognition system. Whenever this predicate is called in an EP pro- gram, the intention recognition system is employed to check if Ag has intention I , i.e. I is the most likely conceivable intention at that moment. We also provide predicate has_intention(Ag,I,Pr) , stating that agent Ag has intention I with prob- ability Pr . Hence, one can express, for example, the situation where one needs to be more, or less, cautious. One can also generalize to consider the N -best intention recognition approach, that is, to assess whether the intention of the agent is amongst the N most likely intentions. It has been shown that by increasing N , the recognition accuracy is sig- nificantly improved (Armentano & Amandi, 2009; Blaylock & Allen, 2003; Han & Pereira, 2011a). In the sequel we draw closer attention to some EP constructs, illustrating with several examples how to take into account intentions of other agents for enhancement of decision ...
Context 3
... BN for recognizing Fox’s intention is depicted in the Figure 2. The initial possible intentions of Fox that Crow comes up with are: Food- i(F) , Please- i(P) and Territory- i(T) . The facts that might give rise to those intentions are how friendly the Fox is ( Friendly_fox ) and how hungry he is ( Hungry_fox ). These figure in the first layer of the BN as the causes/reasons of the intention nodes. Currently, there is only one ob- servation, which is, Fox praised Crow ( Praised ). In this work, Bayesian Network inference will be performed using P-log, a probabilistic logic system, described in the next section. This will not only allow us to effectively represent the causal relations present a BN for intention recog- nition, the logic-based implementation of P-log will also allow us to make an easy integration with the EP system. The P-log system in its original form (Baral, et al., 2009) uses Answer Set Programming (ASP) as a tool for computing all stable models (Baral, 2003; Gelfond & Lifschitz, 1993) of the logical part of P-log. Although ASP has proven a useful paradigm for solving a variety of combinatorial problems, its non-relevance property (Castro, Swift, & Warren, 2007) makes the P-log system sometimes computationally redundant. A new implementation of P-log (Han, et al., 2008; Han, Carroline, & Damasio, 2009), which we deploy in this work, uses the XASP package of XSB Prolog (XSB, 2009) for interfacing with Smodels (Niemela & Simons, 1997), an answer set solver. The power of ASP allows the representation of both classical and default negation, to produce 2-valued models. Moreover, using XSB as the underlying processing platform enables collect- ing the relevant abducibles for a query, obtained by need with top-down search. Furthermore, XSB permits to embed arbitrary Prolog code for recursive definitions. Consequently, it allows more expressive queries not supported in the original version, such as meta-queries (proba- bilistic built-in predicates can be used as usual XSB predicates, thus allowing the full power of probabilistic reasoning in XSB) and queries in the form of any XSB predicate expression (Han, et al., 2008). In addition, the tabling mechanism of XSB (Swift, 1999) significantly improves the performance of the system. In general, a P-log program Π consists of a sorted signature, declarations, a regular part, a set of random selection rules, a probabilistic informa- tion part, and a set of observations and actions. Sorted Signature and Declaration: The sorted signature Σ of Π contains a set of constant symbols and term-building function symbols, which are used to form terms in the usual way. Addition- ally, the signature contains a collection of special function symbols called attributes. Attribute terms are expressions of the form a(t) , where a is an attribute and t is a vector of terms of the sorts required by a. A literal is an atomic expression, p , or its explicit negation, neg_p . The declaration part of a P-log program can be defined as a collection of sorts and sort declarations of attributes. A sort c can be defined by listing all the elements c= {x 1 ,...,x m } or by specifying the range of values c= {L..U} , where L and U are the integer lower bound and upper bound of the sort c . Attribute a with domain c 1 ×... × c n and range c is represented as follows: If attribute a has no domain parameter, we simply write a: c 0 . The range of attribute a is denoted by range(a) . Regular Part: This part of a P-log program consists of a collection of XSB Prolog rules, facts and integrity constraints (IC) formed using literals of Σ. An IC is encoded as a XSB rule with the false literal in the head. Random Selection Rule: This is a rule for at- tribute a having the form: This means that the attribute instance a(t) is random if the conditions in Body are satisfied. The DynamicRange allows us to restrict the default range for random attributes. The RandomName is a syntactic mechanism used to link random attributes to the corresponding probabilities. A constant full can be used in DynamicRange to signal that the dynamic range is equal to range(a) . Probabilistic Information: Information about probabilities of random attribute instances a(t) taking a particular value y is given by probability atoms (or simply pa-atoms) which have the fol- lowing form: meaning that if the Body were true, and the value of a(t) were selected by a rule named RandomName , then Body would cause a(t) = y with probability A/B . Note that the probability of an atom a(t,y) will be directly assigned if the corresponding pa/3 atom is the head of some pa -rule with a true body. To define probabilities of the remaining atoms we assume that, by default, all values of a given attribute, which are not assigned a probability, are equally likely. Observations and Actions: These are, respec- tively, statements of the forms obs(l) and do(l) , where l is a literal. Observations obs(a(t,y)) are used to record the outcomes y of random events a(t) , i.e. random attributes and attributes dependent on them. Statement do(a(t,y)) indicates a(t) = y is enforced as the result of a deliberate action. In an EP program, P-log code is embedded by putting it between two reserved keywords beginPlog and endPlog . In P-log, probabilistic information can be obtained using the XSB Prolog built-in predicate pr/2 (Han, et al., 2008). Its first argument is the query, the probability of which is needed to compute. The second argument captures the result. Thus, probabilistic information can be easily embedded by using pr/2 like a usual Pro- log predicate, in any constructs of EP programs, including active goals, preferences, and integrity constraints. What is more, since P-log (Han, et al., 2008) allows us to code Prolog probabilistic meta- predicates (Prolog predicates that depend on pr/2 predicates), we also can directly use probabilistic meta-information in EP programs. We will illus- trate those features with several examples ...
Citations
... In both cases, we experimented with populations with different proportions of diverse strategies in order to calculate, in particular, what is the minimum fraction of individuals capable of intention recognition for cooperation to emerge, invade, prevail, and persist. It is noteworthy that intention recognition techniques have been studied actively in AI for several decades [27][28][29][30][31], with several applications such as for improving human-computer interactions, assisting living, moral reasoning, and team work [32,33]. In most of these applications the agents engage in repeated interactions with each other. ...
The mechanisms of emergence and evolution of collective behaviours in dynamical Multi-Agent Systems (MAS) of multiple interacting agents, with diverse behavioral strategies in co-presence, have been undergoing mathematical study via Evolutionary Game Theory (EGT). Their systematic study also resorts to agent-based modelling and simulation (ABM) techniques, thus enabling the study of aforesaid mechanisms under a variety of conditions, parameters, and alternative virtual games. This paper summarises some main research directions and challenges tackled in our group, using methods from EGT and ABM. These range from the introduction of cognitive and emotional mechanisms into agents’ implementation in an evolving MAS, to the cost-efficient interference for promoting prosocial behaviours in complex networks, to the regulation and governance of AI safety development ecology, and to the equilibrium analysis of random evolutionary multi-player games. This brief aims to sensitize the reader to EGT based issues, results and prospects, which are accruing in importance for the modeling of minds with machines and the engineering of prosocial behaviours in dynamical MAS, with impact on our understanding of the emergence and stability of collective behaviours. In all cases, important open problems in MAS research as viewed or prioritised by the group are described.
... Such sensitivity to the thought processes of others who may consider cheating or deception as an option involves a further capacity to recognize intentions. Consistent with Kant's observations, our research confirms that intention recognition plays a crucial role in moderating social interactions, even when any given intention is not carried out [30,31,51]. ...
We present a summary of research that we have conducted employing AI to better understand human morality. This summary adumbrates theoretical fundamentals and considers how to regulate development of powerful new AI technologies. The latter research aim is benevolent AI, with fair distribution of benefits associated with the development of these and related technologies, avoiding disparities of power and wealth due to unregulated competition. Our approach avoids statistical models employed in other approaches to solve moral dilemmas, because these are “blind” to natural constraints on moral agents, and risk perpetuating mistakes. Instead, our approach employs, for instance, psychologically realistic counterfactual reasoning in group dynamics. The present paper reviews studies involving factors fundamental to human moral motivation, including egoism vs. altruism, commitment vs. defaulting, guilt vs. non-guilt, apology plus forgiveness, counterfactual collaboration, among other factors fundamental in the motivation of moral action. These being basic elements in most moral systems, our studies deliver generalizable conclusions that inform efforts to achieve greater sustainability and global benefit, regardless of cultural specificities in constituents.
... Many other results of the current literature directly or indirectly related to the naturalization of logic need be quoted, such as recent AI oriented research on counterfactual reasoning [15,56,64,68]; moral reasoning [37,67,74,75]; mutual debugging and argumenting [65,66]; objecting [62,63]; preferring [61]; forgetting [3]; updating [2,4,38]; intention recognition and decision making [31,33]. Also, interesting studies related to the evolutionary game theory concerning emergent population norms and emergent cooperative behavior morals represent a new promising area for the naturalization of the logic of agents embedded in populations and groups, and certainly points out central issues which help to go beyond the expressive rigidity of the mainstream received logical tradition [3,30,32,[34][35][36]. ...
... as well as purely semantic processes that lead to the identification of the array -high level vision" [71, p. 189]. 30 On the basis of this distinction it seems plausible -as Fodor contends -to think there is a substantial amount of information in perception which is theory-neutral. However, also a certain degree of theoryladenness is justifiable, which can be seen at work for instance in the case of so-called "perceptual learning". ...
... The top-down process takes advantage of descending pathways that send active information out from a central point and play a part in selectively "listening" to the environment, involving relevant motor aspects (indeed action is fundamental to calibrating perception). The role of hearing in the perception of space is central, complementing multichannel visual information with samples of the acoustic field picked up by the ears: cues to location of source by means of interaural intensity, difference and distance according to cues like loudness are two clear examples of 30 A full treatment of the problem of perception both from a psychological and neural perspective is available in the recent [72]. ...
I will analyse some properties of abduction that are essential from a logical standpoint. When dealing with the so-called
‘inferential problem’, I will opt for the more general concepts of input and output instead of those of premisses and conclusions,
and show that in this framework two consequences can be derived that help clarify basic logical aspects of abductive reasoning:
(i) it is more natural to accept the ‘multimodal’ and ‘context-dependent’ character of the inferences involved, (ii) inferences
are not merely conceived of in the terms of the process leading to the ‘generation of an output’ or to the proof of it, as
in the traditional and standard view of deductive proofs, but rather, from this perspective abductive inferences can be seen
as related to logical processes in which input and output fail to hold each other in an expected relation, with the solution
involving the modification of inputs, not that of outputs. I will also describe that if we wish to naturalize the logic of
the abductive processes and its special consequence relation, we should refer to the following main aspects: ‘optimization
of situatedness’, ‘maximization of changeability’ of both input and output, and high ‘information-sensitiveness’.
... Many other results of the current literature directly or indirectly related to the naturalization of logic need be quoted, such as recent AI oriented research on counterfactual reasoning [15,56,64,68]; moral reasoning [37,67,74,75]; mutual debugging and argumenting [65,66]; objecting [62,63]; preferring [61]; forgetting [3]; updating [2,4,38]; intention recognition and decision making [31,33]. Also, interesting studies related to the evolutionary game theory concerning emergent population norms and emergent cooperative behavior morals represent a new promising area for the naturalization of the logic of agents embedded in populations and groups, and certainly points out central issues which help to go beyond the expressive rigidity of the mainstream received logical tradition [3,30,32,[34][35][36]. ...
... as well as purely semantic processes that lead to the identification of the array -high level vision" [71, p. 189]. 30 On the basis of this distinction it seems plausible -as Fodor contends -to think there is a substantial amount of information in perception which is theory-neutral. However, also a certain degree of theoryladenness is justifiable, which can be seen at work for instance in the case of so-called "perceptual learning". ...
... The top-down process takes advantage of descending pathways that send active information out from a central point and play a part in selectively "listening" to the environment, involving relevant motor aspects (indeed action is fundamental to calibrating perception). The role of hearing in the perception of space is central, complementing multichannel visual information with samples of the acoustic field picked up by the ears: cues to location of source by means of interaural intensity, difference and distance according to cues like loudness are two clear examples of 30 A full treatment of the problem of perception both from a psychological and neural perspective is available in the recent [72]. ...
A complete revision of mainstream logic is an urgent task to be achieved. This revision will be able to bring logic into a creative rapprochement with cognitive science. This can be achieved by trying to do for logic what over forty years ago Quine and others attempted for epistemology. It is necessary to propose a “naturalization” of the logic of human inference. This paper deals with an examination of how the naturalization process might go, together with some indication of what might be achieved by it. To assist the reader in understanding the naturalization of logic I will take advantage of my own research on the concept of abduction, which vindicates the positive cognitive value of the fallacy of the affirming the consequent thanks to the so-called EC-model (Eco-Cognitive model), and of the recent book Errors of Reasoning: Naturalizing the Logic of Inference (2013) [86], by John Woods. While this paper certainly aims at promoting the research program on the naturalization of logic, it also further advocates the placement of abduction in the research programmes of logic, and stresses to what extent our contemporary philosophical and logical tradition is indebted towards Charles Sanders Peirce, a thinker often praised for his productivity but whose quality and importance are too often overlooked.
... During that period, the world is changing, and the initial intention may be changed to a more appropriate one or even abandoned [80, 11, 26] . To act appropriately , especially when beneficial to take into account intentions of others, as for our previously implemented intention-based decision making frame- work [32, 67, 43, 33, 41, 42], an intention recogni-tion method should take into account these changes, and, when necessary, be able to reevaluate the decision making model itself, depending on some time limit. In addition, as new actions are observed, the model should be reconfigurable to incorporate them. ...
Decision making about which are the scrutinized intentions of others, usually called intention reading or intention recognition, is an elementary basic decision making process required as a basis for other higher- level decision making, such as the intention-based decision making which we have set forth in previous work. We present herein a recognition method possessing several features desirable of an elementary process: (i) The method is context-dependent and incremental, enabling progressive construction of a three-layer Bayesian network model as more actions are observed, and in a context-situated manner that relies on a logic programming knowledge base concerning the context; (ii) The Bayesian network is structured from a specific knowledge base of readily specified and readily maintained Bayesian network fragments with simple structures, thereby enabling the efficient acquisition of that knowledge base (engineered either by domain experts or else automatically from a plan corpus); and, (iii) The method addresses the issue of intention change and abandonment, and can appropriately resolve the issue of the recogni- tion of multiple intentions. The several aspects of the method have been experimentally evaluated in applications and achieving definite success, using the Linux plan corpus and the so-called IPD plan corpora, which are playing sequences generated by game playing strategies needing to be recognized, in the iterated Prisoner’s Dilemma. One other application concerns variations of Elder Care in the context of Ambient Intelligence.
... Intention recognition can be found abundantly in many kinds of interactions and communications, not only in human but also many other species [101] . The knowl-edge about intention of others in a situation could enable to plan in advance, either to secure a successful cooperation or to deal with potential hostile behaviours [106, 83, 29, 27]. Given the advantage of knowing the intentions of others and the abundance of intention recognition among different species, it is clear that intention recognition should be taken into account when studying or modeling collective behavior . ...
... Contrary to other approaches dealing with the integration of (direct or indirect ) information about the past in individual decisions, e.g. in [57, 69, 110, 109], intention recognition is performed using a Bayesian Network (BN) model. BNs have proven to be one of the most successful approaches for intention recognition [12, 100, 78, 21, 29]. Their flexibility for representing probabilistic dependencies as well as causal relations, and the efficiency of inference methods have made them an extremely powerful tool for problem solving under uncertainty [73, 74] , and appropriate to deal with several probabilistic as well as causal dependencies occurring in intention recognition. ...
The mechanisms of emergence and evolution of cooperation in populations of abstract individuals with diverse behavioural strategies in co-presence have been undergoing mathematical study via Evolutionary Game Theory, inspired in part on Evolutionary Psychology. Their systematic study resorts as well to implementation and simulation techniques, thus enabling the study of aforesaid mechanisms under a variety of conditions, parameters, and alternative virtual games. The theoretical and experimental results have continually been surprising, rewarding, and promising.
Recently, in our own work we have initiated the introduction, in such groups of individuals, of cognitive abilities inspired on techniques and theories of Artificial Intelligence, namely those pertaining to both Intention Recognition and to Commitment (separately and jointly), encompassing errors in decision-making and communication noise. As a result, both the emergence and stability of cooperation become reinforced comparatively to the absence of such cognitive abilities. This holds separately for Intention Recognition and for Commitment, and even more when they are engaged jointly.
The present paper aims to sensitize the reader to these Evolutionary Game Theory based studies and issues, which are accruing in importance for the modelling of minds with machines, with impact on our understanding of the evolution of mutual tolerance and cooperation. In doing so, it also provides a coherent bird's-eye view of our own varied recent work, whose more technical details and results are spread throughout a number of well recognized publishing venues, and to which we refer the reader for a fuller support of our claims where felt necessary.
... In order to take into account the intentions of other agents in decision making processes, we integrated into EP a previously and separately implemented, but also LP-based, intention recognition system [35,65,67]. The obtained integrated system can perform intention-based decision making [36,37]. It takes into account recognized intentions of other agents within different constructs of the decision making system, notably intention-triggering goals (e.g. ...
... The system has been applied for providing appropriate assistance for elderly people in the Ambient Intelligence domain [32,33,66] (e.g. upon recognizing the intention of an elderly person staying alone in his apartment, the system derives suggestions on how to achieve his intentions appropriately, taken into account his profiles and different aspects of the living environment); for deriving morally acceptable decisions in moral dilemmas [37,42] (it is known that a key factor in legal and moral judgments is actual intention, which for instance can distinguish murder from manslaughter [43,97]). Furthermore, our recent work has paid much attention to the problem of intention-based decision making in large-scale multi-agent systems [38][39][40][41]. ...
... Furthermore, coffee is preferred to teawhen both options are still available after considering all other constraints -provided the elder is sleepy (fifth line). For further details, and more complete and extended examples, see [37,66]. ...
Intention recognition is the process of becoming aware of the intentions of other agents, inferring them through observed actions or effects on the environment. Intention recognition enables pro-activeness, in cooperating or promoting cooperation, and in pre-empting danger. Technically, intention recognition can be performed incrementally as you go along, which amounts to learning. Intention recognition can also use past experience from a database of past interactions, not necessarily with the same agent. Bayesian Networks (BN) can be employed to dynamically summarize general statistical evidence, furnishing heuristic information to link with the situation specific information, about which logical reasoning can take place, and decisions made on actions to be performed, possibly involving actions to obtain new observations. This situated reasoning feeds into the BN to tune it, and back again into the logic component.
In this article, we provide a review bearing on the state-of-the-art work on intention and plan recognition, which includes a comparison with our recent research, where we address a number of important issues of intention recognition.
We also argue for an integrative approach to intention-based decision-making that uses a combination of Logic Programming and Bayesian Networks.
The mechanisms of emergence and evolution of collective behaviours in dynamical Multi-Agent Systems (MAS) of multiple interacting agents, with diverse behavioral strategies in co-presence, have been undergoing mathematical study via Evolutionary Game Theory (EGT). Their systematic study also resorts to agent-based modelling and simulation (ABM) techniques, thus enabling the study of aforesaid mechanisms under a variety of conditions, parameters, and alternative virtual games. This paper summarises some main research directions and challenges tackled in our group, using methods from EGT and ABM. These range from the introduction of cognitive and emotional mechanisms into agents' implementation in an evolving MAS, to the cost-efficient interference for promoting prosocial behaviours in complex networks, to the regulation and governance of AI safety development ecology, and to the equilibrium analysis of random evolutionary multi-player games. This brief aims to sensitize the reader to EGT based issues, results and prospects, which are accruing in importance for the modeling of minds with machines and the engineering of prosocial behaviours in dynamical MAS, with impact on our understanding of the emergence and stability of collective behaviours. In all cases, important open problems in MAS research as viewed or prioritised by the group are described.
Having addressed the prerequisite issues for a justified and contextualized computational morality, the absence of radically new problems resulting from the co-presence of agents of different nature, and addressed the difficulties inherent in the creation of moral algorithms, it is time to present the research we have conducted. The latter considers both the very aspects of programming, as the need for protocols regulating competition among companies or countries. Its aim revolves around a benevolent AI, contributing to the fair distribution of the benefits of development, and attempting to block the tendency towards the concentration of wealth and power. Our approach denounces and avoids the statistical models used to solve moral dilemmas, because they are “blind” and risk perpetuating mistakes. Thus, we use an approach where counterfactual reasoning plays a fundamental role and, considering morality primarily a matter of groups, we present conclusions from studies involving the pairs egoism/altruism; collaboration/competition; acknowledgment of error/apology. These are the basic elements of most moral systems, and studies make it possible to draw generalizable and programmable conclusions in order to attain group sustainability and greater global benefit, regardless of their constituents.
Machine ethics is a sprouting interdisciplinary field of enquiry arising from the need of imbuing autonomous agents with some capacity for moral decision-making. Its overall results are not only important for equipping agents with a capacity for moral judgment, but also for helping better understand morality, through the creation and testing of computational models of ethics theories. Computer models have become well defined, eminently observable in their dynamics, and can be transformed incrementally in expeditious ways. We address, in work reported and surveyed here, the emergence and evolution of cooperation in the collective realm. We discuss how our own research with Evolutionary Game Theory (EGT) modelling and experimentation leads to important insights for machine ethics, such as the design of moral machines, multi-agent systems, and contractual algorithms, plus their potential application in human settings too.
