Is Free Will a Necessary Ingredient of Machine ... - Riccardo Manzotti

Metadata of the book and chapters that will be visualized online Book series name

Advances in Experimental Medicine and Biology

Book title

From Brains to Systems

Book subtitle

Brain-Inspired Cognitive Systems 2010

Book copyright year

2011

Book copyright holder

Springer Science+Business Media, LLC

Chapter 15 Chapter title

Machine Free Will: Is Free Will a Necessary Ingredient of Machine Consciousness?

Corresponding Author

Family name

Manzotti

Particle Given Name

Riccardo

Suffix

Abstract

Division

Institute of Communication and Behaviour

Organization

IULM University

Address

Via Carlo Bo, 8, 20143, Milan, Italy

E-mail

[email protected]

Sooner or later, machine consciousness will have to address the elusive notion of free will either to dismiss it or to produce a machine implementation. It is unclear whether freedom and consciousness are independent aspects of the human mind or by-product of the same underlying structure. Here, the relevant literature is reviewed focusing on the connection between determinism and freedom—usually explored by compatibilists. Eventually, a tentative model for machine free will is outlined.

1 2

AUTHOR’S PROOF

3 4

Chapter 15

Machine Free Will: Is Free Will a Necessary Ingredient of Machine Consciousness?

5 6 7

Riccardo Manzotti

8 9

10 11 12 13 14 15 16 17

Abstract Sooner or later, machine consciousness will have to address the elusive notion of free will either to dismiss it or to produce a machine implementation. It is unclear whether freedom and consciousness are independent aspects of the human mind or by-product of the same underlying structure. Here, the relevant literature is reviewed focusing on the connection between determinism and freedom—usually explored by compatibilists. Eventually, a tentative model for machine free will is outlined.

18 19

For a time I thought of the problem of the freedom of the will as the most suitable Gordian knot; but in the end I opted for the concept of the mind (Ryle 1970 [31], p. 2)

20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45

Free will might be a necessary ingredient to design autonomous machines. Furthermore, as it has been said for consciousness, the attempt at emulating free will in machines might pave the way to its understanding. Between consciousness and free-will there are complex relationships that did not receive enough attention in recent cognitive literature probably due to demanding ontological and philosophical issues. The recent debate on free will often stresses the relation with consciousness notwithstanding the skepticism of a few authors. Martin Heisenberg remarks that we need not be conscious of our decisionmaking to be free. What matters is that actions are self-generated. Conscious awareness may help improve behavior, but it does not necessarily do so. Why should an action become free from one moment to the next simply because we reflect upon it? [14, p. 164]. Yet, the target of this skepticism seems to be the enabling role of consciousness rather than some broader and deeper connection between consciousness and free will. And it is difficult to criticize Heisenberg’s view although, it must be stressed that he is arguing probably against a conceptual straw man—has anyone really maintained that consciousness is enough to the freedom of a choice? And what kind of consciousness? Heisenberg’s criticism is fueled by a reflective model of consciousness seen as higher order thought. This is by no means the only possibility. Although self determination is a key ingredient of free will, without a more precise understanding of consciousness final conclusions are prematurely drawn. Is it justified to rule out deep relations between consciousness and freedom? Aren’t our intuition endorsing some kind of identity between a free choice and a conscious one? Have one ever taken a free choice while acting automatically and unconsciously? Is it conceivable a subject at the same time conscious and lacking any free will, whatever it is?

46 47 48 49 50

R. Manzotti () Institute of Communication and Behaviour, IULM University, Via Carlo Bo, 8, 20143 Milan, Italy e-mail: [email protected]

51

C. Hernández et al. (eds.), From Brains to Systems, Advances in Experimental Medicine and Biology 718, DOI 10.1007/978-1-4614-0164-3_15, © Springer Science+Business Media, LLC 2011 Book ID: 193000_1_En, Chapter ID: 15, Date: 2011-05-31, Proof No: 1

1

2 52 53

AUTHOR’S PROOF

54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79

R. Manzotti

If freedom is nothing but a name given to a certain level of autonomy and control achieved by cognitive systems? Or does it imply some deeper structural requirement? Since the beginning of western philosophy, a subject has been conceived as a free agent. An action which is not free is not even considered as an action but rather as a mechanical effect of previous causes. Action-hood entails freedom. Yet, very little of this intuitions were of some use in AI and cognitive science in general. It is true that, for a while, free will has been dismissed from scientific debate since it seemed to require some extraingredient as to the physical description of nature. In this respect, freedom and consciousness were held in similar dubious reputation by the scientific community. Furthermore, freedom is akin to consciousness as to presence/absence in animals. It was long maintained that animals were not conscious since they were not provided of soul. Similarly, it was stated that they were not free for the same metaphysical shortcomings. More liberal reasoning suggests that since animals posses some kind of consciousness so they must be free, at least to a certain extent. There is plenty of evidence that animal behavior cannot be reduced to totally fixed responses. A dog seems to show a freer behavior than a fly. Yet, the issue is vague. Is there a critical threshold? Of what kind? Structural, behavioral, or metaphysical? Is such threshold correlated with consciousness? Finally, what about machines? Could they exploit some kind of freedom? In the following—contrary to the layman’s widespread belief that free will is incompatible with determinism—it will be maintained that free will requires a deterministic world. The well known compatibilist stance will be put to good use—namely the view holding that free will is indeed possible only because of determinism. If this view is correct, there is a conceptual and indeed cognitive space for free will in machines. Machines might offer a testbed for theories and models of free will. In particular, this chapter will consider and criticize a model of present—here dubbed “temporal atomism”—which restrains from considering free will as a structural aspect of cognition. Taking advantage of a refutation of temporal atomism, in the last part of the chapter, a few aspects of free behavior—such as integration and polytropism—will be considered in relation to a tentative implementation of free will. It’s fair to warn the reader from the start that if you expect to find a block diagram of freedom you are going to be deluded.

80 81 82 83

15.1 What is Free Will?

84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101

Famously Spinoza wrote that “[M]en believe themselves to be free, because they are conscious of their own actions and are ignorant of the causes by which they are determined” [33]. Yet, if freedom were a matter of pure ignorance, it would be too easy to succeed in replicating it into a machine. According to such a definition, most available artificial agents would indeed be free since they act without the slightest knowledge of the causes of their actions. Obviously something else must be added. When do we consider an agent to be free? A simple and still inadequate answer is that an agent is free if, and only if, it is able to achieve its goals against any external constraints. Yet, this is not enough and it can lead to an apparent freedom as it happens in all those agents where the constraints are internal or simply the action space is either severely limited or fixed. The issue at stake runs much deeper. Is the agent able to choose its goals if its behavior is constrained by previous factors such as programming, environmental stimuli, upbringing, genetic blueprint, and such? According to the philosopher Immanuel Kant, someone’s actions are not free if they are determined by something or someone else. An agent is taken to be free when it seems able to be the only and ultimate cause of its behavior. It is up to the agent whether to do something. Or, at least, this is what it seems. As we have seen, this is not necessarily the case. It could be an illusion due to insufficient causal knowledge. According to Peter Hájícek [12], any acceptable explication of free will has three ingredients: (1) it must entail that an agent might have chosen otherwise; (2) it must explicate the control that

102

Book ID: 193000_1_En, Chapter ID: 15, Date: 2011-05-31, Proof No: 1, UNCORRECTED PROOF

15 103 104

AUTHOR’S PROOF

105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124

Machine Free Will: Is Free Will a Necessary Ingredient of Machine

3

free will requires; and (3) must explicate the “sensibleness” or “rationality” that free will involves. The first requirements thus implies that the agent is able to contemplate and weight more than one course of action and then to single out an outcome on the basis of its own criteria. It is rather obvious from the start that there is some tension in such a formulation—on the one hand free will requires multiple outcomes, on the other hand, it requires the control and the efficacy of the agent’s criteria and rationality. It is plain that control and multiple outcomes are contradictory principles. By and large, currently there is an obvious difference between a human being and a machine. Human behavior is the result of mostly unknown causes that are practically unknowable because of their sheer numbers and their causal role in one’s life. In the case of machines, since they are the result of human design or programming, it is much easier to provide an almost exhaustive causal account. This means that, back to Spinoza, is easier to believe (wrongly) that a human being is freer than a machine, since it is easier to ignore the causes of human behavior. However this is just an epistemic difference and it cannot be used to endorse an ontological divide. Leaving aside Spinoza’s epistemic concerns, it seems we are faced with two possibilities: either an outcome is the unavoidable result of previous cause and thus it is not free, or it is the result of the free choice of the agent and thus it occurs as a result of some special causal power of the agent. Both conditions seem inappropriate for machine freedom. Yet, for identical reasons, they would be inappropriate for human freedom too. This leaves open the possibility that we are either applying unjustified assumptions or drawing unwarranted conclusions. Nevertheless, if human beings and other animals are free, in some practical sense of the word, machines could exploit the same cognitive trait unless humans were powered by some unnatural force alien to the physical world, which seems pretty unlikely.

125 126 127 128

15.2 Free Will Geometry

129 130 131 132 133 134 135 136 137

The discussion on free will often revolves around a set of vaguely defined notions. In this section, a few critical notions which are crucial in the discussion on freedom are briefly listed and discussed: predictability vs. unpredictability, determinism vs. indeterminism, self-determination vs. hetero-determination, internal vs. external. The goal is to get rid of some vagueness before engaging in a more challenging debate. Furthermore, the quick outline of these notions is here relevant insofar as it shows how much the debate of free will is linked with the gap between epistemic and ontological issues. In fact, many of the following dichotomies are rooted in the divide between what is known about an agent’s behavior and the processes leading to that behavior.

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152

Predictability vs. Unpredictability If an agent is free, it is up to it to choose one outcome among many. In principle, there is no way to know in advance what the agent will do. However, in many situations, it is reasonably certain that a certain action will be preferred among many others. If someone buys the winning lottery ticket, what she will do? On the other hand, if you twist the knob to set the timer of a washing machine, assuming that everything is fine, you know perfectly well what is going to happen. To be free is thus to be unpredictable? Up to a point. In fact, Spinoza could be right. The fact that something cannot predicted could be only a matter of epistemic limitation on the beholder’s side. The weather cannot be predict reliably for more than a month in advance, yet as far as we know, the atmosphere is not a free agent. After seven bounces, a billiard ball trajectory is unpredictable. Yet, the ball is not freer because of the chaotic components of its trajectory. On the other hand, it is not unconceivable that a free choice could be rather predictable, although completely free—if such a thing would exist. For instance, once confronted with the possibility of avoiding pain, most humans would gladly do so (not all, though!). Are they less free since they are all the more predictable?

153


4 154 155

AUTHOR’S PROOF

156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176

R. Manzotti

Determinism vs. Indetermism According to absolute determinism, of the kind Pierre Simon Laplace was fond of, every event is completely determined (fixed) by its predecessors. Since nothing comes out of nothing, everything is fixed. Determinism seems an easy view, but it is not [15, 17]. Does determinism imply causation? What kind of connection ties together subsequent events? Further, note that determinism does not entail predictability as shown by chaos theory. A chaotic system can be perfectly determined and yet, since its initial state cannot be measured with perfect accuracy, its future development remains unknown. The comforting picture of a deterministic universe was jeopardized by quantum mechanics since intrinsically casual events were admitted. In other words, although the probability density function of certain events is defined, their individual occurrence is not. For instance, there is a probability of 1/2 that a certain nucleus will change its status but there is no way to cause it to happen at a certain time t . In short, an indeterminate event is an event whose occurrence at a time t is not caused by any previous event or state of affairs. A determinate event is an event whose occurrence is exhausted by previous facts. Indeterminacy entails a rupture with the past, while determinacy seems to guarantee a continuity. Yet. indeterminism is no safe harbor for free will. As it was remarked by William James, “If a free act be a sheer novelty, that comes not from me, the previous me, but ex nihilo, and simply tack itself on to me, how can I, the previous I, be responsible?” [18, p. 53]. Similarly many authors observed that although indeterminism gets rid of unwanted external causes, it also gets rid of the necessary determination that should proceed from the subject towards the choice. Suppose that, whenever an agent has to take a choice between two options, a device throws a quantum dice. Then the agent acts accordingly. Would this behavior be free? It seems unlikely. A free choice is not a random one. A free choice is an act that is at the same time linked with the agent’s past and unconstrained by external causes.

177 178 179 180 181 182 183 184 185 186

Self Determination vs. Hetero-determination Suppose that an agent is constrained in its behavior by external causes. Surely it would fall short of any intuitive notion of freedom. A good example is being programmed by an external designer. If the agent’s goals are set by an external designer, they are not the expression of the agent’s freedom. On the other hand, a free agent is self-determined. Being self-determined is an interesting notion since it stresses the deep relation between determinism and freedom rather than indeterminism and freedom. It also stresses the fact that the notion of the self is crucial and probably prior to that of freedom. The chasm between freedom and necessity closely matches that between agents who are their own cause and agents whose behavior depends on conditions external to them.

187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203

Internal vs. External Causes external to the agent cannot contribute to its freedom. Such causes range from external constraints like chains to more subtle ways to constraint the agent’s behavior such as upbringing, education, genetic background, nature and nurture. Once Stephen Jay Gould wrote that “If we are programmed to be what we are, then these traits are ineluctable.” [10, p. 238]. This is not the place to engage in the nature vs. nurture debate. For what it matters here, both genes and cultural environment and education are just the same: causes which we consider to be external to the agent as such. The issue at stake is that intuitively there seems to be a sharp distinction as to whether some event belongs to the agent. If it does, the event can be considered internal to its structure, otherwise is usually conceived as external. To be internal is not sufficient to contribute to the freedom of the agent, yet to be external is surely sufficient to be alien to its free will. Such sharp distinction is very Cartesian, to say the least. The boundaries of the body are too narrow and too large at the same time. While an external event could be considered legitimately part of an agent’s being, a tumor growing inside the agent’s body would not. On the other hand, a neuron could be external to the agent whether the agent was identical to some neural activity that does not comprise that neuron. On the basis of our notion of what the agent is, a component of the agent is either rejected or accepted.

204


15 205 206

AUTHOR’S PROOF

207


5

To recap, the above sketchy outline of the various views on free will suggests that—whether anything like free will exists—it is determined or, better, self-determined, internal to the agent, intertwined with its individual history and selfhood. Is anything like that available in a machine?

208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254

15.3 Temporal Atomism As to free will, the two main views are compatibilism and libertarianism. According to compatibilism, the agent is determined and freedom is a reality that has to be explained in terms of integration, self-determination, consciousness, and other cognitive features. The opposite of compatibilism is incompabilism which holds that a deterministic world is incompatible with freedom. This entails that if the world were deterministic, there would be no freedom. Alternatively, if freedom does indeed exist, it would meant that the world is not completely deterministic. Incompatibilists who accept free will and deny determinism are called libertarians [1, 5, 16, 20]. From a libertarian standpoint, which is closer to the layman’s view of freedom, the kind of liberty suggested by compatibilism is nothing but a “wretched subterfuge”, as the philosopher Immanuel Kant was fond of saying. Many distinguished authors defended this view suspecting that the alternative would throw away the baby with the bath water [19, 21]. On the other hand, from a compatibilist perspective, which has been embraced by recent scholars like Daniel Dennett or Ted Honderich, libertarianism is nothing but a flight of fancy trying to cope with the hard facts of physics introducing vague metaphysical principles [6, 8, 17]. It has been observed that “AI depends on a compatibilist view, but having taken it, there is a lot to be learned about the specific forms of free will that can be designed” [30, p. 342]. The reason is rather obvious. Up to know, all machines are made of deterministic components and for good reasons too. A machine behavior has to be as much predictable as possible. Of course, if compatibilism would be true, implementing free will in machines would be easier. Yet this is not a proof as to whether compatibilism is true. According to many authors, a critical aspect of free will is the so-called problem of origination. In other words, they look for a crucial place and time where the choice is originated. As Robert Kane remarked “If there is indeterminacy in free will, on my view, it must come somewhere between the input and the output” [19, p. 27]. That place somewhere between is the special locus where free will allegedly changes the course of events. This is rather plain in Benjamin Libet’s words “The initiation of the freely voluntary act appears to begin in the brain unconsciously, well before the person consciously knows he wants to act?” [24, p. 49]. Once more there is the expectation that free will has to originate into a narrow spatio-temporal locus. The notion of free will is thus framed in terms of temporal atomism. By this term, I mean the view that the act of free will must be squeezed into a temporal atom of no width. At some point in the chain, there must have been an act of origination of a new causal chain. In this way, libertarians can defend the pristine causal role of free will. However, it could also lead to the refutation of free will as a physical possibility [34]. Given a causal chain leading to an action, it has been often claimed that free will must intervene at a given point by a voluntary act. It is as if there were a perfectly determined chain of events and then, somewhere and somewhen along the line, some unexpected event happens and possibly changes the course of events. This unexpected event is the alleged free choice of the agent. Given these (unwarranted) premises, temporal atomism is necessary in order to allow the origination of free will. No matter what has happened before (even internally), when the agent chooses, nothing is going to determinate its future outcome. “If we are responsible . . . then we have a prerogative which some would attribute only to God: each of us, when we act, is a prime mover unmoved. In doing what we do, we cause certain events to happen, and nothing—or no one—causes us to cause those events to happen” [4, p. 32].

255


6 256 257

AUTHOR’S PROOF

258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279

R. Manzotti

However both the doctrine of origination and the model of temporal atomism suggests a mythical atomism of the free will act which runs counter any physical model of cognitive processes. In fact, any physical processes requires time to take place. Complex cognitive processes require a lot of time— usually in the order of hundreds of milliseconds. Yet, this should not come as a surprise. Neither should we look for some mythical temporal atom in which the free will occurs squeezed between otherwise determined causal chains. Yet, most empirical research on free will seem to assume that free will, if it exists, has to be located in a temporal instant [13, 24, 32]. Whenever this is not confirmed, it is taken to be a refutation of free will. However, it is not. Rather it is a refutation of the assumed temporal atomism. Consider the classic argument by Libet [22]—since the neural activity underpinning an action begins some time before the conscious choice, the autonomy of the will is challenged. However, it is to be expected that, whatever neural activity endorse cognitive processes is to be spread in time. Cognitive and neural processes are not instantaneous. According to Libet “What we found, in short, was that the brain exhibited an initiating process, beginning 550 ms before the free voluntary act; but awareness of the conscious will to perform the act appeared only 150–200 ms before the act. The voluntary process is therefore initiated unconsciously some 400 ms before the subject becomes aware of her will or intention to perform the act” [23, p. 124]. Subsequent empirical evidence is coherent with these initial findings [11, 13, 32]. Yet why ought to be so surprising? There is no reason to expect that a decisional process is not spread over an extended span of time. There is no limit indeed to how much time a cognitive process can take advantage of. Another reason to criticize libertarianism lies in its intrinsic dualism. In fact it suggests that humans (and free agents in general) exploit some intrinsic property which is not shared by the rest of reality. This is highly dubious. From a scientific point of view, compatibilism is a stronger position since it does not require any extra-hypothesis on the nature of reality.

280 281 282 283

15.4 A Model for Machine Free Will

284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305

Are there cognitive architectures suitable to endorse free will? In the above, I argued that if anything like free will exists, there is no reason why should be present only in human beings. Further, free will is likely to be related neither with some mythical origination nor with casual events. Rather, it is to be expected that free will stems out of very complex causal processes akin to those exploited by human beings. However, it is plain that simple deterministic devices are not up to the task. It is as if, very loosely speaking, there is good determinism and bad determinism. Whereas the latter is the kind usually associated with machines and the former is the one linked to our sense of responsibility and free will. Consider a random machine built in such a way that whenever it has to take a decision, it throws a quantum dice which is genuinely casual. Is it free? No, of course not. First, in such a case, to talk of an agent is purely metaphorical. Secondly, as it has been argued above, to be indeterminate is not sufficient to be free. Consider a pure automata like a simple toy. For instance, a cuckoo clock. This time there is no indetermination involved. Once more there is no freedom either. The machine does not qualify in any intuitive sense as a free agent. It does not qualify as an agent, too. Yet its behavior is completely determined by internal causes. However there are at least two shortcomings. On the one hand, the causes are ultimately triggered by original previous causes external to the cuckoo clock such as the clockmaker starting the mechanism. On the other hand, there are no choices since all events follow a rigid course. Is it possible to design and implement a machine that is capable of true choices albeit being ontologically determinate? A first step in the right direction, to our opinion, was suggested by Gary

306


15 307 308

AUTHOR’S PROOF

309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356


7

Drescher [9] who contrasted situation-action machine with the more complex choice machine, in which the individual creation generates its own reasons for doing x or y, by anticipating probable outcomes of various candidate actions and evaluating them in terms of the goals it also represents [9]. Instead of looking for a magic step that introduces a gap in the causal determinacy of events, more complex agents exploit a more structured intertwinement between their history and their structure. Using Drescher’s terminology, a situation-action machine is a repository of sensory-motor responses. This means that the outcome of the machine is not the result of its individual development. The same idea matches the difference between Darwinian and Skinnerian agents outlined by Daniel Dennett [7]. While the former act purely on the basis of fixed input-output patterns, the latter build on top of their own experiences and modify their behavioural structure accordingly. Likewise, Riccardo Manzotti and Vincenzo Tagliasco suggested to distinguish between teleologically fixed and teleologically open agents [29]; the former resemble closely both Drescher’s situationaction and Dennett’s Darwinian/Skinnerian agents, the latter are a further step in the direction of a thoroughly self generated causal structure. A teleologically open agent is an agent that not only is able to learn how to behave to reach fixed goals, but it is also capable of acquiring new goals. Its history becomes part of its teleological causal structure. The common hunch in all these approaches is that the causal structure of an agent is not fully there from the start. Rather the agent develops while more and more causes get entangled together by means of the peculiar physical structure of the agent. There are two ways to interact between an agent and its environment. In the first case, external events trigger responses out of the agent, yet the agent structure remains the same. Consider a laptop. Its code does not change because of the user and the data it processes. There is no agent there, since there is no development, no causal space for an agent to grow around the seeds of the external stimuli. Consider now a human being. In such a case, during development, external events trigger complex changes in the causal structure of the agent. The external world nurtures and causes the causal structure itself. Is there a disguised dualism here? Is the agent contrasted with its causal structure? Not necessarily, the two are the same. An agent is its causal structure. It seems that there are two roles for causations and only the case in which causation changes the causal structure is relevant for free will. Or at least, this is what emerges out of the available literature. This is not the place to discuss in details what is meant by ‘causal structure’. Assume that it could be something like the capability to store and manipulate internal representations, or the existence of a global workspace, or the capability to develop new goals, or something entirely unknown at the present. This author is agnostic as to what is meant by ‘causal structure’. Whatever it is, it is clear that a standard program does not modify it, while a human being does. Recently, Dennett wrote that, “It is commonly supposed that in a deterministic world, there are no real options, only apparent options. This is false” [8, p. 25]. I do agree. But how can a deterministic machine account for choosing among real options? The question cuts to the crux of the problem. The proposal is that a free agent is capable of real choices. What is a real choice? It is neither a random outcome, nor the result of causes external to the agent. A choice is the outcome of a set of causes that are constitutive of the agent. Is there circularity here? Only if the agent were defined as a system capable of choices. Of course, the burden of the above definition lies on the capability of outlining the nature of the agent. Before embarking on a discussion as to the nature of real choices, it must be stressed that the above definition allows a gradual notion of freedom. In fact, freedom does not arise out of a special component. Rather, free will depends on the complexity of the agent considered. In fact, the freedom associated with each agent’s action depends on how much that action is the result of the agent’s self. For the sake of the argument, let’s assume here that the agent’s self is not a metaphysical entity but rather a quantifiable aspect of a cognitive agent such as the total memory, or the total number of sensory-motor contingencies, or the Tononi’s integrated information, or whatever [35]. If this were

357


8 358 359

AUTHOR’S PROOF

360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407

R. Manzotti

possible, a free choice would depend on how much that particular choice is the expression of the agent individuality (more on this in the following). To recap, literature suggests that there is a progression from situation-action agents to more freedom-oriented agents and that such a progression corresponds to an increased entanglement between individual history and causal structure. This approach allows avoiding the determinism vs indeterminism dilemma. It draws on a different crucial property—namely being free depends on being part of the agent. Since the existence of an agent is a matter of degree, the freedom associated with an action could vary. Contrary to the duchess of Devonshire’s opinion, freedom could indeed be a matter of degree. A minimum of formalization is going to help. Suppose to have an agent and the events {ei } occurring in its surroundings. Suppose that these events can be divided in three possible categories: events {ei∗ } that are not causally influent with the agents (for instance, radio-waves for a worm), events {ei∗∗ } that are causally processed by the agent but that not alter its causal structure (for instance keystrokes on a laptop), and events {ei∗∗∗ } that are going to be embedded in the causal structure of the agent either as representations or as goals (for instance, the first glance at Juliet by Romeo). The third and last category could indeed be proposed as a rough approximation of the kind of events that partake to the causal structure of the agent. In slightly more philosophical words, the events {ei∗∗∗ } are those that have a counterfactual relation with the causal structure of the agent. The agent is here modeled as a collection of casual relations A = {ri |ri = cj → ek } such that they were introduced in the past by some triggering events. In this way, hard-wired casual relations are put aside—the agent is the result of the individual history. Given the agent A, there are three possible cases 1. If any e∗ occurs, nothing happens 2. If any e∗∗ occurs, ∃j ∗ |cj ∗ → e; in other words, because of the agent A, the event produces some outcome. Yet the causal structure of A remains unaltered. 3. If any e∗∗∗ occurs, either one or more new causal relations rk are added to the causal structure of A, or one or more existing causal relation ri are modified, or both. In the above formalization what is interesting is that all is a matter of degree. It is even possible to suggest a way to weight how much an incoming event is part of the subject. The basic idea, surely to be further refined, is that only the incoming ei∗∗∗ belong to the subject and that they belong all the more if they affect either existing or new causal relations constituting the agent. Considering the ratio between the number of the causal relations affected by an event e and the total number of causal relations constituting the agent, a very rough index of belongingness could be estimated. Ideally, an event that is going to change completely the agent’s casual structure, is going to affect all its causal relations with the world. Consider a Paul-on-the-road-to-Damascus kind of event in one’s life. After an experience like that, all is different. Conversely, an isolated and not very important event is going to have a much more local effect. So much for the past events concocting the causal structure of the agent, but what about the choices the agent has to take? Let us skip the question as to whether a free choice could be originated by totally endogenous and unpredicted causes such as is the case when all of a sudden one wants to do something totally unforeseen. A more common case of free will exercise is to take choices in response to incoming stimuli. You see a barking dog: do you stand or run away? Whenever you have to answer to an incoming stimulus, you have to take a choice. How many of your past experiences and thus of embedded causal relations are involved? Keep in mind that the agent is not made of any hardwired input-output patterns. The agent, according to the above definition, is made of what had been gobbled up during its individual history. In order to take a choice, how much of the agent structure has to be put into good use? Suppose that none of it, in fact, is used. It could be. For instance, the agent’s action could be the result of an instinct and thus of a hard-wired reflex like when you abruptly withdraw your hand from a spider on

408


15 409 410

AUTHOR’S PROOF

411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444


9

a wall. Would that be a free choice? Hardly. Suppose, on the contrary, that putting a wedding ring on someone’s finger is the result of a well-pondered balance among most of what constituted one’s life up to that crucial moment. Would that be a free choice? I hope so and, seriously speaking, it looks like. Intuitively, it seems that the more a choice is the result of one’s previous causal structure, the more it is an expression of freedom. Then a free choice would be any outcome which is the result of a relevant portion of the agent’s causal structures. In turn, the agent’s existence is gradually constituted by the increase in the number of causal relations embedded in the individual history. In practice, if there is no individual history, there is no agent. If there is no agent, there is no freedom. In turn, if there are no causal relations of the kind outlined above, there is neither agent nor freedom. So far so good, you may say, but clearly many questions remain unanswered. Against this model of free will in agents and machines, it may be objected that a few crucial things are blatantly lacking. I would like to mention a few. • The whole model might seem somehow circular since it depends on how you define the agent. This is partially true since the agent’s definition is based on the occurrence of counterfactual casual relations of a definite kind. It doesn’t always happen that the occurrence of an event allows for the further occurrence of future events of the same kind. Neither has it happened often that a stimulus changed the perceiver’s structure in any relevant way. Elsewhere, the connection between ontogenesis and epigenesis has be discussed at length [25–29]. Although, of course, the issue is far from being solved, it is likely a yet to be thoroughly understood aspect of agent development. • There is no computational (nor to speak of experimental) evidence of the efficacy of the presented model. True. Yet this is not an argument against the idea in itself but rather an encouragement to further research. • Where is the place where the choice is up to the agent? This would be the libertarian and Cartesian objection to our proposal. There is no place where the agent exploits any magic power to elude the causal determinism of causes and effects. There is no place where the free choice originates. Yet this is neither a surprise nor a shortcoming. In fact, it was to be expected given our sympathy for compatibilism. A machine is not the be expected to elude physical laws and there is no hope that future understanding of complexity theory or other sophisticated mathematical models will justify the emergence of any radical departure from the present standard physical model of reality. • Would a machine satisfying the presented model be free? The suggested model is a model of free will that could express human free will as well. In this sense, being free is mostly the capability to be self determined. Of course the notion of self determination carries a heavy epistemic debt with the notions of self and determination.

445 446 447 448

15.5 Conclusion

449 450 451 452 453 454 455 456 457 458

According the Greek mythology, Apollo was a single minded god. He was monotropos. In fact, he drove the sun chariot along a fixed trajectory in the sky. And for many good reasons, as the disastrous Phaeton’s attempt showed to all. In contrast, the god Hermes had a devious mind. His mind was polytropos since it was able to pursue many different goals and to choose among them. Hermes mind offers a suitable example of what intuition suggests to be a free mind. His capability of endorsing multiple ends is a key aspect of free will. This is something that has been repeatedly stressed here, although, strictly speaking, it is a necessary but not sufficient condition for free will. In fact, if there are no alternatives, there is no freedom. Yet the simple generation of multiple scenarios is not sufficient to guarantee a free choice.

459


10 460 461

AUTHOR’S PROOF

462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487

R. Manzotti

Another issue that has to be mentioned, however briefly, is the role of temporal integration. Since events are likely to take place at different times, an agent made of a huge number of causal relations is likely to be the product of a relevant span of time. This is an example of temporal integration where the causal structure of the agent is the result of events spread along a temporal interval. In turn, any outcome of such an agent is the result of temporal integration of many otherwise separate events. It should come as no surprise that many of the relevant conditions we have outlined for freedom have been already mentioned in previous works on machine consciousness [2, 3, 29]. For instance, polytropism is very close to teleologically openness. Similarly, the kind of causal relation expressing when an event belongs to an agent is tightly tied with previous models of machine consciousness. This is to be expected and considered moderately encouraging. In fact, consciousness and free will share many aspects both from an empirical and a theoretical view. Consider the gap between automatic responses and conscious ones and how close it matches the gap between automatic and free action. It is well known that most automatic sensory-motor reflexes do not require any awareness. In contrast, consciousness steps in whenever an individual, meditated, and original choice has to be taken—that is a free one. It is also well known that the more a behavior becomes automated (because of training or repetition) the more it fades from consciousness thereby being subtracted from conscious and free control. Consider the scene in the movie “Burn after reading” when a startled Harry Pfaffer (George Clooney) finds unexpectedly Chad Feldheimer (Brad Pitt) in his closet. Since Harry had a previous training to shoot unconsciously in an emergence, he shoots at Chad. Then he excuses his action since his shooting was not free being the mandatory outcome of his training. Many more examples could be made. Many issues that have not been adequately developed here such as temporal integration, polytropism, automatic vs conscious responses, unity, and agency, are indeed shared between the issue of consciousness and that of free will either in humans and in machines. This is likely not to be fortuitous. Although the only philosophical account of free will that can be adopted by AI—namely compatibilism—is still far from offering an established and detailed blue print for machine free will, it is remarkable to note that machine consciousness and freedom can be seen as the attempt at finding a more efficient way to situate an agent in an unpredictable and largely unknown environment.

488 489 490

Acknowledgements I wish to thank Antonio Chella for his support and encouragement on various challenging and demanding topics in AI and Cognitive Sciences as well as for his many observations on an earlier version of this paper.

491 492 493

References

494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509

1. Atmanspacher, H., Bishop, R. (eds.): Between Chance and Choice. Interdisciplinary Perspectives on Determinism. Imprint Academic, Exeter (2002) 2. Chella, A., Manzotti, R. (eds.): Artificial Consciousness. Imprint Academic, Exeter (2007) 3. Chella, A., Manzotti, R.: Machine consciousness: a manifesto for robotics. Int. J. Mach. Conscious. 1(1), 33–51 (2009) 4. Chisholm, R.M.: Freedom and action. In: Lehrer, K. (ed.) Freedom and Determinism, pp. 11–44. Random House, New York (1966) 5. De Caro, M.: Libero Arbitrio: Una Introduzione. Laterza, Bari (2004) 6. Dennett, D.C.: Elbow Room. The Varieties of Free Will Worth Wanting. MIT Press, Cambridge (1984) 7. Dennett, D.C.: Kinds of Minds: Toward an Understanding of Consciousness. Science Masters, 1st edn. Basic Books, New York (1996) 8. Dennett, D.C.: Freedom Evolves. Penguin Books, London (2003) 9. Drescher, G.: Made-Up Minds: A Constructivist Approach to Artificial Intelligence. MIT Press, Cambridge (1991) 10. Gould, S.J.: Ever Since Darwin. Norton, New York (1978) 11. Haggard, P.: Voluntary action and conscious awareness. Nat. Neurosci. 5(4), 382–385 (2002) 12. Hájícek, P.: Free will as relative freedom with conscious component. Conscious. Cogn. 18, 103–109 (2009) 13. Haynes, J.-D., Sakai, K., Rees, G., Gilbert, S., Frith, C., Passingham, R.E.: Reading hidden intentions in the human brain. Curr. Biol. 17, 323–328 (2007)

510


511 512

AUTHOR’S PROOF

513 514 515 516 517 518 519 520

15


14. 15. 16. 17. 18. 19. 20. 21. 22.

Heisenberg, M.: Is Free Will an Illusion? Nature 459, 164–165 (2009) Honderich, T.: The Consequences of Determinism, vol 2. Oxford University Press, Oxford (1988) Honderich, T.: How Free Are You? Oxford University Press, Oxford (2003) Honderich, T.: On Determinism and Freedom. Edinburgh University Press, Edinburgh (2005) James, W.: Pragmatism: A New Name for Some Old Ways of Thinking. Dover, New York (1907) Kane, R.: The Significance of Free Will. Oxford University Press, Oxford (1996) Kane, R. (ed.): The Oxford Handbook of Free Will. Oxford University Press, New York (2001) Kane, R.: A Contemporary Introduction to Free Will. Oxford University Press, New York (2005) Libet, B.: Unconscious cerebral initiative and the role of conscious will in volutary action. Behav. Brain Sci. VIII, 529–566 (1985) Libet, B.: Mind Time. The Temporal Factor in Consciousness. Harward University Press, Cambridge (2004) Libet, B., Freeman, A., Sutherland, K.: The Volitional Brain: Towards a Neuroscience of Free Will. Imprint Academic, Thorverton (1999) Manzotti, R.: A process based architecture for an artificial conscious being. In: Seibt, J. (ed.) Process Theories: Crossdisciplinary Studies in Dynamic Categories, pp. 285–312. Kluwer Academic, Dordrecht (2003) Manzotti, R.: From artificial intelligence to artificial consciousness. In: Chella, A., Manzotti, R. (eds.) Artificial Consciousness, pp. 174–190. Imprint Academic, London (2007a) Manzotti, R.: Towards artificial consciousness. Comput. Philos. Newsl. 07(1), 12–15 (2007b) Manzotti, R.: From consciousness to machine consciousness. Proc. Addresses Am. Philos. Assoc. 82(1), 54 (2008) Manzotti, R., Tagliasco, V.: From “behaviour-based” robots to “motivations-based” robots. Robot. Auton. Syst. 51(2–3), 175–190 (2005) McCarthy, J.: Free will—even for robots. J. Exp. Theor. Artif. Intell. 12(3), 341–352 (2000) Ryle, G.: Autobiographical. In: Wood, O.P., Pitchers, G. (eds.) Ryle: A Collection of Essays, pp. 1–15. Double Day, Garden City (1970) Soon, C.S., Brass, M., Heinze, H.-J., Haynes, J.-D.: Unconscious determinants of free decisions in the human brain. Nat. Neurosci. 11, 543–545 (2008) Spinoza, B.: The Ethics (Ethica Ordine Geometrica Demonstrata). Dodo Press, New York (1664) Strawson, G.: Free will. In: Craig, E.M. (ed.) Routledge Encyclopedia of Philosophy. Routledge, London (1998/2004) Tononi, G.: An information integration theory of consciousness. BMC Neurosci. 5(42), 1–22 (2004)

23. 24.

521 522

25.

523 524 525 526 527

26. 27. 28. 29.

528 529

30. 31.

530 531

32.

532 533 534 535

33. 34. 35.

536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561


11