Artificial Homeostatic System: A Novel Approach

Share Embed


Descrição do Produto

Artificial Homeostatic System: A Novel Approach Patrícia Vargas1, Renan Moioli1, Leandro N. de Castro2, Jon Timmis3, Mark Neal4, and Fernando J. Von Zuben1 1 DCA/FEEC/Unicamp - Brazil {pvargas, moioli, vonzuben}@dca.fee.unicamp.br 2 Unisantos - Brazil [email protected] 3 University of Kent - UK [email protected] 4 University of Wales - UK [email protected]

Abstract. Many researchers are developing frameworks inspired by natural, especially biological, systems to solve complex real-world problems. This work extends previous work in the field of biologically inspired computing, proposing an artificial endocrine system for autonomous robot navigation. Having intrinsic self-organizing behaviour, the novel artificial endocrine system can be applied to a wide range of problems, particularly those that involve decision making under changing environmental conditions, such as autonomous robot navigation. This work draws on “embodied cognitive science”, including the study of intelligence, adaptivity, homeostasis, and the dynamic aspects of cognition, in order to help lay down fundamental principles and techniques for a novel approach to more biologically plausible artificial homeostatic systems. Results from using the artificial endocrine system to control a simulated robot are presented.

1 Introduction In previous work, Timmis and Neal [22] presented a model for an artificial endocrinesystem (AES) as a module of a broader conceptual framework including artificial neural networks (ANN) and artificial immune systems (AIS), with the ultimate goal of developing an artificial homeostatic system (AHS). The AHS (e.g., a mobile robot) is capable of autonomously interacting with an unknown and changing environment while maintaining its internal state (e.g., energy level and integrity) and optimizing some objectives. It is now believed in biology that there is an interconnection and dependence among the immune, nervous and endocrine systems, which is fundamental for cognition, maintenance of the internal state of an organism (called homeostasis), immune-regulation and host defenses [1]. The present work concentrates on the neuro-endocrine interactions and mechanisms in order to create a more biologically plausible artificial homeostatic system. The approach borrows some ideas from “embodied intelligence”, introduced by Rodney Brooks [2][3][4] to synthesize a cognitive system based on coupled dynamics and nonstationary mappings. M. Capcarrere et al. (Eds.): ECAL 2005, LNAI 3630, pp. 754 – 764, 2005. © Springer-Verlag Berlin Heidelberg 2005

Artificial Homeostatic System: A Novel Approach

755

The paper is organized as follows: Section 2 gives a brief description of the fundamentals of the nervous and endocrine systems, followed by some key interactions between the two systems and the biological mechanisms that motivated this work. Section 3 revisits some aspects of the embodied cognitive science related to robot autonomous navigation. In Section 4 the previous model of the artificial homeostatic system is presented, and Section 5 introduces the novel artificial homeostatic system. Some preliminary simulations are presented in Section 6. Discussions and future work compose Section 7.

2 Nervous and Endocrine Systems Interactions The nervous system (NS) is primarily responsible for the reception of stimuli, by detecting changes in the internal and external environments, and for processing and transmitting the nerve impulses as appropriate responses to those changes [14]. The endocrine system can be viewed as a system of glands [20] that works with the nervous system in controlling the activity of internal organs and in coordinating the long-range response to external stimuli. The main roles of the endocrine system are to assist the maintenance of homeostasis, growth, differentiation, metabolism, reproduction, and to help the organism to cope with stress [14][15]. All these tasks are accomplished by the hormones, which are chemical substances produced, stored and secreted by the components of the endocrine system, including a group of glands, specialized cells, body tissues and organs. One of the most important neuro-endocrine interactions happens within the hypothalamus and the pituitary gland (or hypophysis). The hypothalamus is a region in the brain beneath the thalamus. It consists of many aggregations of nerve cells and its main function is to control release of pituitary hormones. The hypothalamus is responsible for the integration of many basic behavioural patterns, involving the correlation of neural and endocrine functions. Its neurons are also affected by a variety of hormones and other circulating chemicals [10]. In fact, due to this interaction, in the last three decades the hypothalamus has been often referred to as the “endocrine hypothalamus” [9][20]. The pituitary gland is located at the base of the brain. The hypothalamus controls the release of hormones from the pituitary that will in turn control other target organs, and also other target endocrine glands. This hypothalamus-pituitary interaction is controlled by feedback mechanisms. There exists a positive feedback when the production and release of hormones is excitatory. Nonetheless, normally these physiological functions are under a negative feedback mechanism regulation, in which the hormone production and release is inhibitory [10]. Feedback is carried out by three mechanisms: a sensor that will sense the controlled variables under supervision, a reference point, and an error signal. Release of some hypothalamic hormones is ruled by external and internal neural inputs (short loops), and also by long feedback loops involving remote organs and external metabolic processes [9]. The important point is that the release of hormones influences nervous activity (thus cognition, motor control, etc.), whilst nervous activity influences endocrine function (thus growth, metabolism, etc.), in a semi-closed control loop: internal

756

P. Vargas et al.

processes are allowed to regulate themselves whilst external environmental factors can also regulate and control the system. In the novel homeostatic system to be developed here, an artificial endocrine system will be adopted to aid an artificial neural network in the process of robot autonomous navigation.

3 Robot Autonomous Navigation – A Cognitive Challenge Attempts to understand the mind and its operation have been ongoing since the ancient Greeks philosophers [21]. This study has lead to the development of cognitive science. Simply put, cognitive science is the interdisciplinary study of mind and intelligence [19]. This in turn has lead to the appearance of a novel field of research at the end of the 1980’s, called “embodied cognitive science”, also known as “behaviourbased robotics”. The new term was based on the ideas of “embodied intelligence” coined by Rodney Brooks [2][3][4]. For Brooks, the only way to understand intelligence is by giving it a body, i.e. by turning it into an embodied system it will be possible to focus upon the interaction between the embodied system and the real world, therefore releasing it from the human interpreter. These ideas are best understood when potential applications for intelligent robots are considered. Sea and submarine prospecting, space exploration, discovery of mines, firefighting, military assistance, and search and rescue services are among the plethora of challenges that might be faced by an autonomous mobile robot. There are many ways of conducting artificial agent experiments productively and systematically, among which there are experiments designed for simulated and real robotic agents. The merits of simulation versus physical embodiment are still under discussion (see Pfeifer and Scheier [19]). In our work both approaches are adopted. There is a real robotic agent, the Khepera II® Robot, and a simulated robot agent, using the WSU Khepera Simulator [18]. Among the methods for tackling robot autonomous navigation problems, it is worth highlighting evolutionary approaches [6][7][8][17]. Evolutionary theory proposes that the brain has evolved to control behaviour in order to ensure our survival [19]. Additionally, it is agreed that intelligence manifests itself in behaviour and thus we must understand behaviour before we can completely understand intelligence and therefore create embodied intelligence. Toward this goal, another extremely important concept, which can be considered behaviour–based, is adaptivity, i.e. the ability to adapt to a continuously changing and unpredictable environment. In fact, there is a direct relation between intelligence and adaptivity [19]. During adaptation, some variables need to be kept within certain pre-determined bounds, either by evolutionary changes, physiological reactions, sensory adjustment, or by simply learning novel behaviours. This definition of adaptivity has to do with the concept of homeostasis, a term first coined by Ashby in 1960 [19]. Within the limits controlled by homeostatic processes, the organism or the artificial agent can function and stay alive in a “viability zone”. While trying to design artificially homeostatic (and self-organizing) systems, this work proposes a novel biologically plausible artificial homeostatic system based on previous work for autonomous navigation [22].

Artificial Homeostatic System: A Novel Approach

757

4 Artificial Homeostatic System – Previous Work In previous work, Timmis and Neal [22] presented a model for an artificial endocrine system (AES), which would be part of a broader conceptual framework including artificial neural networks (ANN) and artificial immune systems (AIS) (Figure 1). The system was based on the mammalian body and its mechanisms for the maintenance of homeostasis, where the authors have chosen an intermediate level of granularity.

Fig. 1. Artificial homeostatic system overview (adapted from Timmis and Neal [22])

The AES was described as a system that employed controlling hormones. In the present paper, only the AES and ANN interactions will be discussed (the entire homeostatic system is discussed in greater detail in Timmis & Neal [22]). The ANN proposed uses a standard error backpropagation-learning algorithm to train a multi-layer perceptron (MLP) neural network [11]. Initially there is no explicit interaction between the ANN and the AES. The AES provides a medium-term regulatory control mechanism for the behaviour of the system. It consists of gland cells that secrete hormones stored using a pool mechanism in response to external stimuli:

rg = α

nx

g



i=0

xi

(1)

where rg is the quantity of hormone released by gland g; αg is the rate at which hormones are released by gland g; xi is the i-th input to gland g; and nx is the number of inputs to gland g. Given that cg(t) is the hormone concentration in gland g at a time t, then the variation in concentration obeys (where β < 1 is a decay constant): c g (t + 1) = c g (t ) × β

(2)

Gland cells secrete and record the concentration of hormones present in the system and use it to moderate the strength of reaction. Each gland cell secretes a specific hormone, represented by a simple bit-string. The hormone levels would affect the input weights in the ANN, i.e. the recorded hormone level would affect each input weight on a particular neuron as follows:

758

P. Vargas et al. nx

ng

i =0

j =0

u = ∑ wi xi ∏ C j S ij M ij

(3)

where u is the internal activation of the neuron, wi is the weight for the i-th input xi; n is the number of weights; ng is the number of glands in the system; Cj is the concentration of hormone j; Sij is the sensitivity of the connection from receptor i to hormone j; and Mij is the match between receptor i and hormone j.

5 Artificial Homeostatic System: A Novel Approach Towards the goal of designing a more biologically plausible model and a system with a greater potential to promote artificial homeostasis, we focused on mimicking some mechanisms of the endocrine system used to control the concentration level of hormones within the artificial agent (a simulated robot in our study). A novel artificial endocrine system is composed of three main modules: hormone level repository (HL), hormone production controller (HPC), and endocrine gland (G) (Figure 2). The hormone level repository has a record of the level of hormone in the organism; the hormone production controller is responsible for controlling the production of hormones in response to variations in the internal state of the organism and external stimulation; and the endocrine gland receives inputs from the HPC, being responsible for producing and secreting hormones when required. Note that, in such a system, any variation in the internal and external states may promote or suppress the activity of the nervous (ANN) and endocrine (AES) systems. For instance, the variation of the internal state of the organism as a result of hormone production may act as a feedback mechanism to the hormone production itself, resulting in the release of inhibitory hormones or in the cessation of hormone production. Internal inputs Artificial Endocrine System

HL

HPC

G

ANN

External inputs

Fig. 2. The main components of the new AES and their interaction with the environment

The system dynamics is founded on some of the main biological mechanisms of homeostasis, particularly positive and negative feedback mechanisms of the endocrine system. The HPC module sends excitatory signals, which work as a positive feedback to the gland G, which in turn starts to produce and release hormone (without implementing the pool mechanism previously proposed [22]), thus increasing the hormone level. The level of hormone will in turn alter the internal state by driving neural network actions upon the environment. By sensing inhibitory signals that promote

Artificial Homeostatic System: A Novel Approach

759

negative feedback from the internal state, the HPC module ceases the production of excitatory signals (positive feedback) until once again it senses specific changes in the internal state. 5.1 System Dynamics The internal state (IS) of the system described in this paper is modeling a seeking behaviour, i.e. it drops to zero or null state when the robot reaches the target. By drawing an analogy to human drives and desires, the target can be governed by many types of possible behaviors, such as, desire to eat: target = food; desire to charge: target = base; desire to wander: target = no collision. The internal state of the artificial agent will depend on the level of external stimulus (ES) and also on the hormone level (HL) present within the artificial organism at instant t. If the ES and HL are above certain pre-determined thresholds (λ and ω, respectively), then the internal state is equal to zero; that is, the level of “desire” falls down to zero. This happens because there are enough external stimuli present and there are enough hormones to trigger the behaviour. Otherwise, the internal state will increase at a pre-determined rate β until it reaches a pre-defined maximum level Max(IS): If (ES ≥ λ ) and (HL ≥ ω ) then IS = 0 else IS(t + 1) = IS(t ) + β (Max(IS) − IS(t ))

(Rule 1)

The external stimulus (ES) depends on the proximity of the artificial agent to the targets. Analogous to the human body sensitivity to external stimuli, this distance can be sensed by the artificial agent using its sensor inputs. This information will be made available not only to the artificial neural network but also to the artificial endocrine system. The hormone production controller (HPC) effectively “selects” in a completely sub-symbolic way the behaviour to be exhibited by the robot. The HPC mimics the hypothalamus and thus senses both the external environment and the internal state, and thus triggers the hormone production and release by the gland G. This triggering may cause the hormone concentration level to increase and therefore to stimulate its target cells (the neurons) to perform a certain task. While this task is not accomplished, the HPC will keep on producing and releasing hormones, thus maintaining a suitable hormone level. When the task is accomplished, i.e. when the HPC receives a negative feedback signal, it ceases the production and release of hormone. Based on this mechanism, Rule 2 synthesizes the control of hormone production and release: If IS≥θ then HP = 0

else

HP(t + 1) = (100− %ES) ×α(Max(HL) − HL(t))

(Rule 2)

where θ is the target threshold of the internal state IS; HP is the hormone production; ES is the external stimulus; α is the scaling factor; HL is the hormone level; and t is the time index. If the internal state IS is greater than or equal to a target threshold θ, then hormone will be produced at a rate that will depend upon the level of the external stimulus received and the level of hormone already present within the artificial organism.

760

P. Vargas et al.

Otherwise, if the internal state IS is less than a target threshold θ, then hormone production will cease. The hormone level represents the amount of hormone stimulating the neural network (ANN). The hormone level will undergo a constant updating in its value due to its internal half-life measure [8] and the amount of hormone produced: HL(t + 1) = HL(t ) × e −1/ T + HP(t )

(4)

where T is the hormone half-life.

6 Preliminary Experiments In order to better clarify and assess the performance of the new AHS three preliminary experiments were conducted. Experiments I and II were designed for a simulated robotic agent using the WSU Khepera Robot Simulator [18] (Figure 3) and Experiment III was designed for a real robotic agent using the Khepera II Robot® [13]. The rationale behind these experiments is twofold: first, to show the systems’ adaptivity through its ability to cope with internal and external changes; and, second, to confirm the systems’ ability to adapt to a dynamical environment, by presenting the phenomenon of biological cyclic behaviour synchronized with the amount of resources available via homeostatic control. 6.1 Experiments with a Simulated Robotic Agent The simulated robot first learned two separate tasks: to avoid collision and to detect a light source (here associated with a food source). Both tasks were learned via a standard error backpropagation-learning algorithm used to train two separate multilayer perceptron (MLP) neural networks [11]. The input-output training data is composed of samples from a diverse set of relevant navigation conditions. After training, the robot was introduced into the arena with walls (obstacles) and a light source (food source) in the middle (Figure 3).

Fig. 3. Experiment I: Sequence of steps performed by the robot. The robot leaves the top right corner moving towards the light source (in the middle of the arena).

Artificial Homeostatic System: A Novel Approach

761

In Experiment I, once the robot begins to navigate, its behaviour was solely controlled by the AES, which was designed to manage its internal state “desire to eat” (here, eat means recharge) employing the endocrine mechanisms described previously. The robot’s light sensor values ranged from 50 to 500 depicting the presence of external stimulus (high values mean lack of food). Note that the robot reaches the target, fulfils its “desire to eat” and then moves away from the target towards the left wall. All other values were printed from the system behaviour/reaction to this information. Figure 4a shows the hormone and the internal state levels and Figure 4b shows the external stimuli during 55,000 iterations or navigation steps. The hormone and internal state values ranged from 0 to 100 units at most. The internal state depicted in Figure 4a refers to the need of energy (desire to eat). The highest level of hormone production added to the proximity to the target caused the robot to fulfil its “desire to eat” four times during the simulation. This confirms the influence of the hormone level over the robot’s autonomous behaviour. The parameter values adopted in this simulation were: β = 0.0001; α= 1.8; T = 500.0; λ= 150; ω= 90.0; and θ = 75.0, and were defined empirically.

(a)

(b)

Fig. 4. Experiment I: (a) Hormone and internal state levels.(b) Output of the light sensor along navigation.

Experiment II explored a scenario where the robot actions were influenced by two external stimuli, the distance to the target and a varying, cyclical availability of food, meaning that the higher the food quantity (FQ), the greater the amount of food available. This experiment was divided into two distinct simulations: A and B. In Simulation A, the FQ was created based on a smooth sine curve function and its value had an influence over the parameter of the internal state (Rule 1), providing a self-adaptive behaviour for the robot in a way that it tends to eat only when the FQ value is higher than 50 (in a 0 to 100 scale) (Figure 5a). This adaptability accounts for the cyclic behaviour observed when the robot’s drive to eat synchronized with the availability of food. In Simulation B, the FQ was created based on a stepwise sine curve function (to facilitate a future simulation designed for a real robotic agent) and its value had also an influence over the β parameter of the robot’s internal state (Figure 5b). For instance, the FQ could be implemented by an automatic discrete electronic apparatus for

762

P. Vargas et al.

the control and adjustment of the light source blinking speed. The faster the blinking speed, the greater the quantity of food available.

(a)

(b)

Fig. 5. (a)Experiment IIA: synchronous behaviour for a sinusoidal stimulus (b) Experiment IIB: synchronous behaviour for a cyclical stepwise stimulus

6.2

Experiments with a Real Robotic Agent

In Experiment III, the concepts of embodied intelligence were incorporated into a real robotic agent. The main idea of this experiment is to use the same control system developed for the simulated robotic agent in a real robotic agent. There was just a fine tuning of the input sensors

Fig. 6. Experiment III(a): A complete trajectory of the real Khepera robot, in an environment surrounded by walls, with a light source in the middle (b): An extended simulation from a different starting point.

Figure 6a shows a complete trajectory in an environment surrounded by walls, with a light source in the middle. The robot initiates the navigation searching for the nearest wall and starts to follow it. When its internal state level exceeds a predetermined limit, the artificial endocrine system determines the increase of the hormone level. This will cause the robot to follow the light (to recharge its battery). As soon as the robot reaches the light, the hormone level starts to decrease and the robot switches

Artificial Homeostatic System: A Novel Approach

763

back to the wall-follower behaviour. Figure 6b illustrates the same idea, but shows a higher number of behaviour changes from a different starting point.

7 Conclusion This paper presented a biologically inspired artificial homeostatic module of an artificial endocrine system based on a previous work for robot autonomous navigation. It also contributed with fundamental principles and concepts for the novel approach towards the goal of creating artificially homeostatic (and self-organizing) systems. The experiments conducted support the adaptability capacity of the artificial homeostatic system (simulation and experiments with real robots) through its ability to cope with internal and external changes, and also the ability to adapt to a dynamical environment by presenting a cyclic behaviour synchronized with the resources available. We made the robot react in synchrony with the environment without any internal world model and only using sub-symbolic representation. The results presented also show the basic operation of the control loop generally associated with homeostasis, where internal processes and external environmental factors are allowed to regulate the system. The ultimate perspective is that the system proposed constitutes part of a new form of implementing embedded cognitive science. Comparative analysis of the current proposal and alternative Bio-inspired mechanisms to endow an artificial organism with autonomy is being considered as a further step of the research.

References 1. Besendovsky, H. O. & del Rey, A. (1996). Immune-Neuro-Endocrine Interactions: Facts and Hypotheses. Endocrine Reviews, 17(1), pp. 64-102. 2. Brooks, R. & Breazeal, C. (in press). Embodied intelligence. Cambridge, MA, MIT Press. 3. Brooks, R. (1991a). “Intelligence without representation”. AI, 47, pp. 139-160. 4. Brooks, R. (1991b). “Intelligence without reason”, in J. Mylopoulos and R. Reiter (Eds.) Proceedings of the 12th Int. Joint Conference on AI, Morgan Kaufmann. 5. de Castro, L. N. & Timmis, J. (2002). “The Immune System in Context with Other Biological Systems”, Chapter 5 in Artificial Immune Systems: A New Computational Intelligence Approach, Springer-Verlag: London. 6. Dorigo, M., Colombetti, M. (1997). Robot Shaping: An Experiment in Behavior Engineering (Intelligent Robotics and Autonomous Agents), MIT Press. 7. Floreano, D., Mondada F. (1995a). “Evolution of homing navigation in a real mobile robot”, IEEE Transactions on System, Man and Cybernetics. 8. Floreano, D., Mondada F. (1995b). “From evolution of innate behaviours to evolution of learning in robotics agents”, Technical Report R95.06I, Laboratory of Microcomputing, Swiss Federal Institute of Technology at Lausanne. 9. Greger, R. and Windhorst, U. (1996), (eds.) “Comprehensive Human Physiology” (Vols. 1 and 2). Springer. 10. Guyton, A. C. & Hall, J. E. (1996). Textbook of Medical Physiology, 9th ed., W. B. Saunders Company, Philadelphia.

764

P. Vargas et al.

11. Haykin, S. (1999). Neural Networks: A Comprehensive Foundation, 2nd ., Prentice Hall. 12. Klopf, A. H. (1982). The Hedonistic Neuron: a Theory of Memory, Learning, and Intelligence, Washington: Hemisphere. 13. K-Team S.A.. (2003). URL: http://www.k-team.com 14. McClintic,J. R. (1975). Basic Anatomy and Physiology of the Human Body,J.Wiley & Sons 15. McClintic, J. R. (1985). Physiology of the Human Body, 3rd ed., John Wiley & Sons 16. Neal M. and Timmis J. (2003). Timidity: A Useful Mechanism for Robot Control? Informatica, Vol. 27 No. 4, pp. 197-204. 17. Nolfi S. and Floreano, D. (2000). Evolutionary Robotics: The Biology, Intelligence, and Technology of Self-Organizing Machines, The MIT Press. 18. Perretta, S. J. and Gallagher, J.C. (2003). The Java Khepera Simulator from the Wright State University, Ohio, USA, http://ehrg.cs.wright.edu/ksim/ksim.html. 19. Pfeifer, R. and Scheier, C. (1999). Understanding Intelligence, MIT Press. 20. Purves W. K., Heller, H. C, Orians, G. H. and Sadava, D. (2001). Life: The Science of Biology, 6th Edition, IE-Macmillan UK. 21. Thagard, P. (1996). Mind: Introduction to Cognitive Science, The MIT Press, USA. 22. Timmis, J. and Neal, M. (2004). “Once More Unto the Breach: Towards Artificial Homeostasis”, in L. N. de Castro and F. .J. Von Zuben, Recent Developments in Biologically Inspired Computing, Idea Group Inc., Chapter XIV, pp. 340-366.

Lihat lebih banyak...

Comentários

Copyright © 2017 DADOSPDF Inc.