Autonomous machine having an adaptive controller

Abstract

The self-regulating machine is configured to provide an observable measurement of the mechanical energy consumed by the machine relative to the mechanical energy stored by the machine, such that the decision-making of the control system of the machine can be based on the actual energy stress of the machine. The self-regulating machine has a control system with predictive models of the internal and external environments. Both predictive models are based on the same set of information representing the common energy basis of the machine. The set of information includes (i) a plurality of interaction signals indicating the direct interactions of the machine, and (ii) a plurality of non-interaction signals indicating information available to the machine without consuming energy. The plurality of non-interaction signals includes emulated signals necessary to ensure that the predictive models of the internal and external environments are based on equivalent parameter sets.

Description

Technical Field

[0001] The present invention relates to a machine including an offline rechargeable power source (e.g., a rechargeable battery, etc.) and a control system configured to autonomously maintain a positive energy state of the machine by monitoring the interaction between the machine and its internal and external environments.

Background Art

[0002] A machine is an assembly of one or more parts intended to obtain some mechanical or other advantage when performing work. An autonomous machine (also referred to herein as an "automaton") typically operates within a defined environment using a control system. An autonomous machine typically includes an offline energy storage capacity and means for recovering the offline energy storage capacity corresponding to the finite capacity of the offline storage and the autonomous function of the machine.

[0003] It is desirable for a machine to limit its energy loss by maximizing its energy efficiency. Particularly in the case of an automaton, maximizing efficiency not only promotes maximizing the period required for the machine to recover its energy storage, but also promotes the case where a functioning automaton must store sufficient energy to facilitate its recovery.

[0004] There are multiple ways to implement a machine controller, and the complexity of the controller to some extent reflects the complexity of the machine it controls. The controller ranges from simple mechanical devices to complex electronic processing means that can also use multiple electromechanical or other transducers as sensors or mechanical actuators.

[0005] A machine controller can employ feedback, and a measured value of the output of the machine functions as a driver for its input. Within the operating range allowed by the mechanical capabilities of the machine, with sufficient and stable gain in its feedback system, the machine can obtain performance mainly limited by its sensing accuracy.

[0006] The machine controller can also employ feedforward, explicitly or otherwise, perhaps by estimating the most likely result from a "Markov chain", using previously acquired knowledge embedded in a model of the machine or its environment, to predict the (nominal) optimal machine action.

[0007] Rather than incorporating a hard encoding of the previously acquired information, an "adaptive" machine controller can utilize its learning ability, where the controller updates its prediction model in response to events and results measured by the machine, thereby "adapting" the model, and thus the machine, to its environment.

[0008] Adaptation model optimization by gradient ascent / descent is usually in line with the "InfoMax principle", where being driven by comprehensive negative feedback, minimizing the prediction error helps to maximize the accuracy, and minimizing the complexity of the predictor helps to maximize its efficiency (in the controller and the resulting actions). However, the convergence provided by the inherent negative feedback of the machine controller is likely to achieve locally optimal actions that do not necessarily coincide with globally optimal results. To enable the search for better non-local solutions, making the necessary adjustments to that convergence via additional positive feedback is known as "wandering".

[0009] More complex adaptive machine controllers can employ so-called "deep learning" methods, in which one or more "hidden layers" establish intermediate correlations between the input and output of the controller. For example, a Helmholtz machine may employ a circuitous "neural network" somewhat similar to the flow of information in the central nervous system [2], where the adaptive "recognition" element models the machine controller input (sensation) and the adaptive "generation" element models the output of the machine (action).

[0010] A superset of control system analysis applicable to machine controllers is the "Free Energy Principle" [3], which describes the minimization of "free energy" in a feedback system [4]. Note, however, that "free energy" is a theoretical measure, not actual thermodynamic energy. Controllers based on the "Free Energy Principle" can generate Bayesian inference networks [5], where posterior (output) predictions are developed from the conditional probabilities of prior predictions, and free energy provides a measure of the "surprise" difference between the controller's predictions and its perceived reality.

[0011] In an ideal inference network, the recognition or generative models are updated to minimize their information redundancy (or reduce deviation from the most efficient representation), such that well-defined modeled objects generate data "clusters" that provide the basis for higher-level representations [6].

[0012] Limited by sensory modalities and their sensitivities, such controllers can identify or infer data clusters from information incidents at the machine's interface, and the data clusters enable the subsequent development of hierarchical models where modeled components are distinguished by a so-called "Markov blanket" [7]. The Markov blanket provides the model with a "scale", where low-level model components are incorporated into high-level model components, and some "value" called "reward" or "utility" (or, complementarily, "loss" or "cost") is assigned, and accordingly the machine controller executes some action determined at that particular scale.

Summary of the Invention

Means for Solving the Problems

[0013] Most generally, the present invention provides an autonomous machine (e.g., a computer-controlled device) configured to provide an observable measurement of the mechanical energy consumed by the machine relative to the mechanical energy stored by the machine, such that decision-making by the control system of the machine can be based on the actual energy stress of the machine.

[0014] In particular, the present invention can provide an autonomous machine having a control system with predictive models of the internal and external environments, both predictive models being based on the same information set representing the common energy basis of the machine. The information set can include (i) a plurality of mutual signals transmitted on a mutual channel, where the plurality of mutual signals indicate direct interactions of the machine ("near-field"), and (ii) a plurality of non-mutual signals transmitted on a non-mutual channel, where the plurality of non-mutual signals indicate information available to the machine without consuming energy ("far-field"), where the plurality of non-mutual signals include emulated signals necessary to ensure that the predictive models of the internal and external environments are based on an equivalent parameter set. The control system can also provide a dedicated parallel path for common-mode signals within the internal environment of the machine. This ensures that signals within the parameter set related to the common mode are orthogonal to the differential-mode signals in both predictive models, thereby enabling the predictive models to distinguish between internal and external power dissipation, and thus providing a true indicator of the energy stress on the machine.

[0015] The present invention can be distinguished from prior art that uses higher-order representations in that the higher-order components of the predictive model are not "separated" from the lower-order components by a Markov blanket. With the architecture described herein, only the error-prone parts (e.g., layers) of the predictive model cause flux in the feedback loop between those models, thereby ensuring that the actual energy basis of the mechanical interactions of the machine permeates the entire relevant machine controller portion and renders the error at the interface of the machine.

[0016] The present invention can be implemented in any autonomous machine, which can maintain itself within its environment through training or other means, where the machine includes an offline energy storage capacity rather than a constantly connected power source, and the machine can recover its offline energy storage by mechanically interacting within its environment.

[0017] Accordingly, according to the present invention, a machine capable of autonomous operation is provided, the machine including a rechargeable offline power supply, a physical interface for the machine to interact with the external environment, the physical interface being operable in differential mode and common mode, configured to provide movement via a differential mode output and to provide movement cancellation of the common mode output, a non-interactive interface configured to passively interact with the external environment via non-interactive inputs (such as sensors), a first control element configured to control the internal and external operations of the machine and to maintain a positive energy state of the machine, and a second control element disposed between the first control element and the physical interface and configured to mediate information exchange therebetween, the second control element being configured to communicate with the physical interface via a first set of mutual information channels and to communicate with the first control element via a second set of mutual information channels, the first set of mutual information channels including an input channel configured to transmit information to the machine from non-interactive inputs and an emulated output channel forming a mutual pair with the input channel, the second control element including a first prediction model configured to predict communication on the first set of mutual information channels and a second prediction model configured to predict communication on the second set of mutual information channels, the first prediction model and the second prediction model both being adaptive models coupled in a feedback configuration to minimize an error energy flux between the first set of mutual information channels and the second set of mutual information channels, the first control element including a common mode regulator configured to maintain a common mode bias at the physical interface and to establish an independent parallel common mode path that appears within the internal environment of the machine.

[0018] The control system architecture defined herein provides certain advantages over the ability of a prediction model to distinguish clusters of information representing various aspects of the internal and external environments that a machine can reason about. In particular, the combination of (i) ensuring that the common mode signal is orthogonal to other signals on both the input and output sides of the second control element, and (ii) using channel emulation to establish fully mutual information channels on both sides of the second control element, means that the prediction model has a common energy basis, where the cluster of information related to the "near-field" portion (related to the mutual interface) can be distinguished from the cluster of information related to the "far-field" portion (related to the non-dissipative interface), and it is possible to identify the cluster of information within the near-field portion related to internal and external effects. This identification ability manifests itself as an improvement in the sensitivity of the prediction model to internal and external influences.

[0019] The rechargeable offline power source can be a battery or other suitable portable power source. In some examples, the power source may include a substrate forming part of the machine, such as a consumable part of the housing or other body of the machine.

[0020] The physical interface may include any suitable structure for exerting a force on the environment to achieve a physical effect. In particular, the mutual interface may be configured to enable the machine to move within the external environment, for example, to access a location where the rechargeable offline power source can be recharged. Thus, the mutual interface establishes an actual energy basis between the first control element and the physical interface. The mutual interface may include, for example, a pair of mutual servo motors. The mutual interface may be configured to provide motion from the differential mode output and motion cancellation from its common mode output. A separate set of mutual elements is required for each dimension in which motion is required.

[0021] The non-reciprocal interface may be any suitable passive sensor for detecting the characteristics of the external environment. The sensor may be an image sensor or scanner configured to detect information about the surroundings of the machine. Alternatively or additionally, it may include environmental sensors for measuring, for example, temperature, humidity, etc. The non-reciprocal interface may also include (nominal) outputs such as a light source. However, it can be understood that such outputs are non-dissipative from the perspective of the mechanical energy delivered to the environment. The non-dissipativity of the non-reciprocal interface means, for example, providing a non-reciprocal information channel that represents the far-field of the machine, and the non-reciprocal information channel is converted into a reciprocal pair by an emulated output channel, and the emulated output channel itself is essentially non-dissipative. When the non-reciprocal interface also includes (nominal) outputs, the first set of reciprocal information channels further includes an emulated input channel that forms a reciprocal pair with the output channel that conveys information to the non-reciprocal interface. In the architecture shown herein, a reciprocal pair is formed for each non-reciprocal input or output. Thus, the machine may have a plurality of non-reciprocal channels, and thus a plurality of corresponding emulated channels, each provided from either a first control element or a physical interface.

[0022] The non-reciprocal channels and their corresponding emulated channels are replicated to a second set of mutual information channels, such that the first prediction model and the second prediction model operate on the same set of information. However, in some cases, the emulated channels do not necessarily need to "pass through" completely between the first control element and the physical interface. Instead, the emulated channels can terminate at a second control element, thereby potentially forming stub channels. One example of a stub channel can be the output from the sequence memory of the first control element. This output can potentially affect future actions but does not need to be input to the physical interface. Another example of a stub channel can be the input from a color detector at the physical interface. This input can potentially form part of the remote field information available to the machine but does not need to be input to the first control element.

[0023] The first control element may be configured to adjust the positive energy state of the machine, for example, using a properly configured feedback loop. The feedback loop for the positive energy state of the machine may include one or more auxiliary feedback loops. For example, the first control element can use a first feedback loop to control the internal environment of the machine and a second feedback loop to cause the operations necessary to restore its off-line power supply. The first control element can be implemented using a control model that associates actions or sequences of actions with results. For example, the model can encode operation instructions that can control the physical interface, enabling the machine to function in its environment. In one example, the control model may include a memory configured to store one or more action sequences, and each action sequence is associated with a result that records the influence of the sequence on the goals of the feedback loop.

[0024] The model may include goals or drivers that function to determine the priorities of actions available in a given scenario. For example, a basic driver may be related to adjusting the positive energy state of the machine. Subsequently, other drivers may be arranged in a cascade, thereby forming a hierarchy that reflects the decrease in their priorities. By arranging a series of drivers in the first control element, one or more drivers can also be made to stop initially until the machine acquires sufficient knowledge to ensure its long-term energy stability.

[0025] The energy state of the machine may be related to its operating state. For example, if the accumulated energy becomes zero, leading to the stoppage of the machine, the machine may be configured to end its operation. The first control element may be configured to control a series of drivers based on this.

[0026] The first control element may further incorporate an adaptive learning module configured to update the model. This model can be adapted, for example, using Pavlovian learning.

[0027] The common mode bias maintained by the first control element may be configured to result in positive internal power dissipation in the interface output. Thus, the common mode regulator may be configured to adjust the internal dissipation via either a feedback or feedforward configuration with respect to some criterion.

[0028] In some embodiments, the internal dissipation from the common mode bias can be used to provide heating to emulate or form a heat-generating machine, where a thermal sensor is utilized to adjust the criterion of the common mode regulator and compensate for changes in the internal temperature.

[0029] The independent parallel common mode paths may be physically defined within the machine in some cases, or may be emulated paths formed within the first control element in other cases.

[0030] To maintain the actual energy basis established by the physical interface of the second control element, both the first set of mutual information channels and the second set of mutual information channels may include a plurality of mutual channel pairs. If the input to the second control element is non-mutual, for example, because it is related to a non-mutual interface, the second control element is configured to emulate the corresponding output to create a mutual pair. Each emulated mutual channel is configured to be nominally non-dissipative.

[0031] The first prediction model is configured to generate an output that includes signals in one or more of the first set of mutual information channels. The second control element may be configured such that predictions formed from that first model (which may represent the external environment) are projected onto all sensed inputs encompassed by the physical interface, even if from a single sensed input. Similarly, the second prediction model is configured to generate an output that includes signals in one or more of the second set of mutual information channels. Thus, the error flux apparent in only a subset of the relevant mutual information channels can still be mapped to the information on all channels.

[0032] If the prediction models are accurate enough and no error is recognized in either model, the second control element operates in a converged state and there is no flow of information exchange through the second control element between the first control element and the physical interface. The advantage of this situation is that the model (and thus the machine) becomes optimally adapted to its environment. Only when an error is recognized in either prediction model does the flow of information flow through the second control element, resulting in the emergence of a local negative feedback loop and being driven to equalize the model impedance. This configuration can be used to adapt the prediction models, resulting in the ability to identify and add new model components during the operation of the machine.

[0033] In one embodiment, the path to convergence in the feedback loop formed within the second control element may be driven in part by the error component on the common mode output.

[0034] At the moment of defining the path to convergence, the control system architecture described herein endows the machine with its own relative energy model, which is rendered in parallel with and indistinguishable from the perceived energy reality that forms the near-field interaction projected within the outer environment of the far-field model. This means that a prediction model can be adapted based on information that distinguishes the near-field and far-field of the machine, where the near-field is the field that is in direct contact with the machine and dissipates the actual mechanical power (either internally or externally). Further, by providing a common mode regulator, it becomes possible to distinguish between the internal power dissipation and the external power dissipation of the machine, thereby providing information indicating the energy stress applied to the machine.

[0035] When the second control element operates in the divergence stage (i.e., when there is an error in one or both of the prediction models), the model error in the second control element is observable at the point of the physical interface or the mutual channel emulation (if not the same). The error can be effectively rendered to the physical interface in an appropriate sensing mode.

[0036] The first prediction model and the second prediction model may each adopt a hierarchical model or a hierarchy model. When configured as described above, all layers below the source of the error are essentially canceled out, and as a result, the energy information rendered to the physical interface is generated by the upper hierarchical layer of the converging model.

[0037] The first and second prediction models, and optionally the control model within the first control element, can be a modular model including a plurality of replaceable units, such as layers within a hierarchical model that can be switched as needed.

[0038] In some embodiments, the path to convergence in the second control element can be adjusted to include a shorter divergence period for overall convergence, i.e., "wandering". For example, in embodiments where the first control element includes a memory of sequences and results, and each result is associated with a positive or negative score for each driver unit, the results can be used to adjust the behavior of the second control element, for example, by restricting the range of Bayesian search to local minima. In another example, the second control element may be adjusted based on the available energy, and that information is available from the first control element. The machine may be configured to suppress wandering until the model in the first control element operates stably.

[0039] The machine may be part of a group of machines having the same control architecture. In some embodiments, the second control element may be configured to introduce perturbations into the first prediction model and / or the second prediction model to introduce variations into the model across the group of machines. In particular, the perturbations may be introduced into the layers of the first or second prediction model when in a converged state. In this scenario, the perturbations may become a permanent part of the learned model.

[0040] The machine can be distributed and arranged on one or more physical sub-components interconnected in a suitable manner, for example, via a wireless network, and the prediction model is established across all physical sub-components and the interfaces therein.

[0041] In one embodiment, the first control element itself may be configured to affect the learning process within the second control element by providing a non-linear response to one or more of the second set of mutual information channels.

[0042] The machine may be operable in both a physical (real) environment and a virtual (e.g., simulated) environment, and references to the "external environment" of the machine herein can be interpreted accordingly.

[0043] The first and second control elements may be implemented as software or firmware executed, for example, on a processor within a machine, as a computer control unit.

[0044] The present invention includes the described aspects and combinations of preferred features, except where such combinations are clearly not permitted or are explicitly avoided.

[0045] Embodiments and experiments for explaining the principles of the present invention will be described below with reference to the accompanying drawings.

Brief Description of the Drawings

[0046]

Figure 1

Figure 2

Figure 3

Figure 4

Modes for Carrying Out the Invention

[0047] Aspects and embodiments of the present invention will be described below with reference to the accompanying drawings. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this specification are incorporated herein by reference.

[0048] The configurations described herein provide two adaptive predictors arranged between the interface of a machine controller and its operating environment. The first predictor predicts events sensed in the environment of the machine, and the second predictor predicts the response of the machine controller.

[0049] The present invention calibrates two predictors through measuring some common mode output orthogonal to a differential mode signal that generates or measures the differential power output of a machine. In the described configuration, an adaptive predictor can be considered to control the (nominally mechanical) impedance presented to a machine interface and a controller.

[0050] The predictor is connected to a local loop so as to minimize the information (energy) flowing through it (nominally done by the approximate simultaneous convergence of separate predictors, although positive feedback may also be utilized to facilitate learning through initial divergence). When the predictor employs a hierarchical memory network (such as a Bayesian hierarchical network), convergence occurs step by step.

[0051] The information flowing through the loop during these steps results from the residual error energy in the prediction. Impedance cancellation is such that a machine controller reacts to these errors (and thus any new information included) in relation to the impact on the energy state (or some function thereof) of the machine (prediction).

[0052] When the predictor is configured to drive "one - to - many" (nominally all) outputs, the machine interface reveals a rendering of that energy itself at the center of its perceived environment projected as an object of energy value to the machine (seen by the machine controller). As a result, its own model is orthogonal to the environment and is always evident during divergence within the loop.

[0053] The present invention further provides a machine including one or more interfaces that may be in different environments, where the advantages of the present invention are realized across all machine interfaces. The present invention can also be configured to accommodate multiple machines (controllers), each machine having its own energy state and a driver for nominally maintaining that state.

[0054] Figure 1 is a schematic diagram showing the functional components of a machine 100 according to an embodiment of the present invention. The machine 100 includes an off-line rechargeable power source 102, a control system 104, and a physical interface 110 that interacts with the external environment 112, for example, by providing the mobility necessary to enable connection to an external energy source for power recharge.

[0055] The control system 104 includes a first control element 106 and a second control element 108, which interact to maintain a positive energy state of the machine through adjustment of the machine and its internal environment 114, as well as control of operation and interaction with the external environment 112.

[0056] The internal environment of a given machine is usually well understood and is generally represented herein by a discrete matrix α. Conversely, the external environment can always change and is defined by a more abstract matrix β.

[0057] In the following description, for the sake of clarity, all indicators referring to time are omitted. Furthermore, since it can be understood that those skilled in the art can implement the following teachings using known techniques for encoding information in the control system 104, it will not be described in detail. In fact, the general machine architecture disclosed herein is suitable for various implementations and applicable to autonomous machines suitable for use in various applications.

[0058] In a general sense, it can be understood that the physical interface 110 of the machine interacts with the external environment 112 in various ways that can be treated as a plurality of interfaces. The output of the machine on these interfaces

Number

Number

[0059] Interfaces can be classified into two categories: (i) mutual interfaces and (ii) non-mutual interfaces. A mutual interface is an interface through which a machine consumes energy in direct interaction with the external environment. In this embodiment, the control system is configured to operate the mutual interface in a common mode and a differential mode.

[0060] In the case of m mutual interfaces, for the differential mode output,

Number

[0061] On the other hand, for the common mode output,

Number

Number

[0062] Therefore, the power consumed by the internal impedance can be expressed as follows:

Number

Number

Number

[0063] The mechanical efficiency when the machine performs some mechanical action can be represented by ηm, where

Number

[0064] A non-reciprocal interface is an interface that transmits information in a way that substantially no energy is lost to the external environment. A non-reciprocal interface includes one or more non-reciprocal outputs and one or more non-reciprocal inputs. A non-reciprocal input can represent, for example, a sensor reading or other passive observation value of the external environment.

[0065] In a general sense, the machine 100 can be understood as having n non-reciprocal interfaces consisting of p non-reciprocal output interfaces and q non-reciprocal input interfaces in total. According to the definition,

Number

Number

[0066] Considering that the machine 100 has a total of σ interfaces (σ = m + n), the following relationship can be expressed for the total power (energy) consumed by the machine.

Number

[0067] The physical interface 110 is the interaction with the external environment on the above-mentioned interface

Number

Number

[0068] From the above, it can be understood that the m mutual channels may be energy-conserving, and as a result,

Number

[0069] Due to energy conservation, it is also possible to make the number of mutual channels less than the number of mutual interfaces. For motion in one spatial dimension, at least two mutual interfaces are required, and one additional interface is required for each additional dimension.

[0070] To simplify the following explanation, for p non-mutual outputs

Number

Number

[0071] More generally, all input channels

Number

Number

Number

Number

Number

Number

[0072] Note that there is power WΞ dissipated in the internal environment and the control system. This is not related to the power consumed for the operation of the machine discussed in this specification, but exists as a component of the differential equation describing the overall energy state E of the machine. Here,

Number

[0073] The control system 104 (which may also be referred to herein as the "controller") includes a first control element 106, the function of which is to adjust the internal environment 114, for example, by using a first feedback loop to control the internal environment of the machine and by using a second feedback loop to cause the operations necessary to restore the offline power supply as needed. For example, the first control element incorporates a positive feedback sufficient to promote energy consumption, such as that required to identify and connect to an energy source necessary for the restoration of power supply, while at the same time providing an overall negative feedback that promotes energy efficiency and maintains a positive energy state of the machine. Thus, the first control element 106 includes the logic necessary to direct the operation of the physical interface 110 in order to achieve the desired purpose, that is, to enable the machine to function in the external environment.

[0074] Expressed in a general sense, the first control element 106 can be understood to interact with the internal environment 114 via a set of j internally facing input channels

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

[0075] Therefore, for the part of the element that controls the operation

Number

Number

[0076] Therefore, considering again that the orthogonal common mode signal and the differential mode signal are added in m mutual channels, the following relationship can be derived.

Number

Number

Number

Number

[0077] As described above, the control system is configured to operate the mutual interface in common mode and differential mode. In this embodiment, the first control element applies a common mode bias configured to bring positive internal power dissipation to the mutual interface. The adjustment of this internal dissipation is made possible through a common mode regulator 116 (for example, a common mode servo amplifier) arranged in parallel with the internal environment 114 and configured to operate in either a feedback configuration or a feedforward configuration with respect to a reference (as an appropriate component of α or β22). The reference may be a permanent energy source partially formed from some positive feedback element and also functions as a basic source of the overall positive energy state of the machine.

[0078] The addition of the common mode regulator helps to maintain the common mode power output W0 and moves the common mode signal away from β11 through the path defined by β22. Defining θ = A + B, where

Number

Mathematics

Mathematics

Mathematics

Mathematics

[0079] p non - mutual outputs

Mathematics

Mathematics

Mathematics

Mathematics

[0080] Due to the action of the common - mode regulator 116,

Mathematics

Mathematics

[0081] The control system 104 further includes a second control element 108 disposed between the first control element 106 and the physical interface 110. The second control element 108 uses two prediction models described below to offload the physical interface from the first control element and vice versa, and mediates between the first control element 106 and the physical interface 110.

[0082] To establish an actual energy basis for operating the prediction model within the second control element 108, the second control element 108 is configured to create an emulated mutual channel for each non-mutual channel it processes. In this way, four types of emulated mutual channels can be created. · An input channel facing the outside

Number

Number

Number

Number

Number

Number

Number

Number

[0083] A common actual energy basis is established within the second control element 108 by the energy conservation link to the external environment within the physical interface 110 described above. To consider the emulated channels, it is useful to rewrite the k outputs and l inputs of the total number σ of interfaces in a more general formulation.

[0084] On the output side of the second control element,

Number

Number

Number

Number

Number

Number

Number

[0085] A similar approach can be adopted on the input side of the second control element.

Number

Number

Number

Number

Number

Number

Number

[0086] The emulation of the mutual channel provides the creation of a quasi-mutual channel that is (nominally) a non-dissipative interface source or load termination as required, that is,

Number

Number

Number

[0087] As described above, the second control element 108 provides two different prediction models that independently model the information presented by the mutual channel

Number

[0088] FIG. 2 is a schematic diagram showing the interrelationship of the prediction models 120, 122 used in the second control element 108 in one embodiment of the present invention. FIG. 2 shows the implementation of the concept in a simplified analog circuit having a single channel for illustrative purposes only.

[0089] On the output side, an external matrix

Number

Number

Number

Number

Number

[0090] On the input side, an internal matrix

Number

Number

Number

Number

Number

[0091] In FIG. 2, for clarity, a pair of operational amplifiers 124, 126 are shown as each receiving a single channel. These elements provide a virtual ground "current" summing node, while two negative converters 130, 132 (e.g., negative impedance converters) act to invert the outputs of models 120, 122 to achieve a cancellation effect. Each of models 120, 122 can be understood as a passive impedance, but represents a time-varying impulse impedance function that acts to modulate the impedance presented to a feedback loop provided within the second control element 108.

[0092] Each of models 120, 122 can be generated using a Bayesian hierarchical network. In general, the models are configured to exhibit a negative impedance with respect to the associated channel or to provide a similar effect. The second control element 108 is configured such that predictions it forms from model 122 of the external environment are nominally projected onto all of the sensed inputs subsumed by the physical interface 110, even if from a single sensed input.

[0093] As shown in FIG. 2, the prediction models 120, 122 are coupled in a manner that, in the absence of the models (i.e., in the absence of the second control element), means that the machine defaults to directly controlling the physical interface 110 by the first control element 106.

[0094] However, by implementing an adaptive learning technique in the second control element 108, the models 120, 122 can adapt to gradually offload the physical interface 110 from the first control element 106, and vice versa. When the models 120, 122 have sufficient fidelity such that the error energy ε is not recognized, the models operate in a converged state as shown in FIG. 3. Here, there is no flux in the coupling of the models 120, 122 for the following reasons.

Number

Number

[0095] When an error is apparent in the models 120, 122 (i.e., when an unknown perturbation has occurred), the coupling flux establishes a feedback loop in the second control element 108 as shown in FIG. 4. The coupling flux drives the models 120, 122 back towards convergence by selecting a more faithful model or extracting (and possibly learning) a new model. Since the error energy ε in both prediction models will always be the same,

Number

Number

[0096] the effects shown in FIGS. 3 and 4 are

Number

Number

Number

Number

[0097] The most efficient models include hierarchical structures, such as the aforementioned Bayesian hierarchical network. In such a configuration, layers are selected and convergence occurs step by step within local feedback loops, so convergence is stepwise. If the requirements for convergence are allowed and sufficient positive feedback is generated by the overall control loop, convergence may also involve further divergence ("wandering") before converging to some non-local equilibrium point.

[0098] In the divergent state (Figure 4), the model error in the second control element can be observed at the physical interface or at the point of cross-channel emulation. The error can be effectively rendered to the physical interface in an appropriate sensing mode. If the second control element employs a hierarchical model or a hierarchical model, all layers below the source of the error are basically canceled out, and as a result, the energy information rendered to the physical interface is at the moment that defines the path to convergence within the local feedback loop of the second control element that occurs when the upper hierarchical layer in the model converges.

[0099] More specifically, during the divergence period in the second control element (i.e., Yε and

Number

[0100] The hierarchical model provides a lower hierarchical section where errors are not recognized and an upper hierarchical section where divergence is apparent. Due to the effective impedance cancellation provided by the prediction model, only the upper hierarchical layer appears in the effective load rendered to the second control element.

[0101] Therefore, the second control element provides "nesting" of errors. More specifically, the energy error identified by the external environment model is adjusted by the requirements determined by each part of the first control element model and may then be rendered to all interfaces within the second control element.

[0102] The model components suitable for selection when assembling the prediction models 120, 122 can be developed as a Markov blanket formed around a cluster of information that can be appropriately described by more efficient means at a higher level. The model components can be formed by adaptation or established in some pre-programmed form.

[0103] In one example, the second control element may be configured to apply a predetermined deterioration term in the first and second prediction models such that they gradually decay over time. In such a manner, only the cross terms that are enhanced by frequent use survive, and thus the no-longer relevant "memory" is discarded.

[0104] Due to the architecture of the control system described above, particularly the provision of the common mode regulator and the cross-channel emulation, the prediction model becomes capable of distinguishing between various clusters of information. In the basic layer of the model, for the convergence iteration step, during use, in the coupling flux within the second control element, there is a cluster of information representing the measured values of the physical energy machine itself and its actions on its energy state, and there is also a cluster of information (or objects) correlated with the mechanical power output of the machine or its prediction (defining the near field of the machine), which can be distinguished from information that is not correlated with the mechanical power output of the machine or its prediction (forming the perceived far field of the machine).

[0105] Regarding the cluster of information defining the near field, through the further identification provided by the common mode servo and the correlation of the power outputs of the common mode and differential modes, relative measurements of internal power dissipation or their predictions, and measurements or their predictions representing external mechanical power dissipation are provided. As a result, a cluster of information representing the machine itself is inferred, and this cluster is distinguished from the far field according to the boundaries inherent in its mechanical interactions, and is subject to an energy stress where a certain action is apparent. The actions include determining the steps for convergence within the second control element and determining the actions by the first control. And, for example, based on the relative energy value of that part of the perceived reality with respect to the machine determined by the iteration of the first control element output, the iteration error is effectively rendered to provide a window that focuses on the entire perceived reality of the machine.

[0106] Note that the rendering of the relative energy measurement values becomes apparent only during the period from when the model error is recognized until convergence is reached. Therefore, the rendering is asynchronous and can occur at discrete instants in real time.

[0107] The first control element may be configured to operate relative to a reference that can indicate a positive energy state of the machine. In one example, the machine can emulate a heat generating machine by configuring the reference to indicate the internal temperature of the machine. For example, the reference may be provided by the output from a common mode regulator having one or more thermometers that sense the internal temperature of the machine. Accordingly, the machine may be configured to control its own heating via modulation of the common mode power output.

[0108] The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, are presented in their specific forms, or from the perspective of means for performing the disclosed functions, or methods or processes for obtaining the disclosed results, and can be used, as necessary, to implement the present invention in various forms by such features individually or in any combination.

[0109] Although the present invention has been described in conjunction with the above exemplary embodiments, many equivalent modifications and variations will be apparent to those skilled in the art when the present disclosure is provided. Therefore, the above-described exemplary embodiments of the present invention are considered to be illustrative and not restrictive. Various changes can be made to the described embodiments without departing from the spirit and scope of the present invention.

[0110] To avoid misunderstanding, the theoretical explanations provided in this specification are provided for the purpose of enhancing the reader's understanding. The inventors do not wish to be bound by any of these theoretical explanations.

[0111] The headings of any sections used in this specification are for purposes of organization only and are not to be construed as limiting the described subject matter.

[0112] Throughout this specification, including the following claims, unless the context requires otherwise, the words "comprise" and "include" and variations thereof (such as "comprises", "comprising" and "including") are to be understood to mean that they include the recited integer or step, or group of integers or steps, but do not exclude other integers or steps, or group of integers or steps.

[0113] Note that, as used in this specification and the appended claims, the singular forms "a", "an", and "the" include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" a particular value and / or to "about" another particular value. When such a range is expressed, another embodiment includes from one particular value and / or to another particular value. Similarly, when values are expressed as approximations by use of the antecedent "about", it will be understood that the particular value forms another embodiment. The term "about" with respect to a numerical value is optional and means, for example, + / - 10%. References [1] Linsker (1988) "Self-organization in a perceptual network". IEEE Computer, 21(3), 105 - 117 [2] Dayan (1995) "The Helmholtz Machine". Neural Comput. 7, 889 - 904 [3] Friston (2006) "A Free Energy Principle for the Brain". J. Physiol. Paris 100, 70 - 87 [4]Feynman(1972) "Statistical Mechanics: A Set of Lectures". Benjamin ISBN°9-7800805-325093 [5]Mackay(2003) "Information Theory, Inference, and Learning Algorithms". Cambridge University Press ISBN 9-780521-642989 [6]Tishby(1999) "The Information Bottleneck Method". 37th Allerton Conference on Communication, Control, and Computing 368-377 [7]Friston(2008) "Hierarchical Models in the Brain". PLOS Comput. Biol. 4(11)

Claims

Claim 1 A machine capable of autonomous operation, wherein the machine comprises a rechargeable offline power supply, and a physical interface for the machine to interact with the external environment, a mutual interface configured to be operable in a differential mode and a common mode, provide motion via a differential mode output, and provide motion cancellation of the common mode output, the physical interface including a non-mutual interface configured to passively interact with the external environment via non-mutual inputs, a first control element configured to control the internal and external operations of the machine and maintain a positive energy state of the machine, a second control element disposed between the first control element and the physical interface and mediating information exchange therebetween, wherein the second control element is configured to communicate with the physical interface via a first set of mutual information channels and communicate with the first control element via a second set of mutual information channels, wherein the first set of mutual information channels includes an input channel configured to transmit information from the non-mutual inputs to the machine and an emulated output channel forming a mutual pair with the input channel, wherein the second control element includes a first prediction model configured to predict communication on the first set of mutual information channels and a second prediction model configured to predict communication on the second set of mutual information channels, wherein both the first prediction model and the second prediction model are adaptive models coupled in a feedback configuration to minimize an error energy flux between the first set of mutual information channels and the second set of mutual information channels, wherein the first control element includes a common mode regulator configured to maintain a common mode bias at the physical interface and establish an independent parallel common mode path within the internal environment of the machine, the machine. Claim 2 The machine according to claim 1, wherein the differential mode output of the mutual interface is configured to move the machine within the external environment, and the mutual interface includes at least one pair of independent mutual elements for each dimension of movement. Claim 3 The machine according to claim 1 or 2, wherein the first control element is configured to adjust a positive energy state of the machine.

4. The first control element is implemented by a control model that encodes instructions that can control the physical interface to enable the machine to perform actions in its environment, the control model being configured to operate based on an internal reference indicating the energy state of the machine, the control model employing a first feedback loop for controlling the internal environment of the machine and a second feedback loop for causing actions necessary to restore the offline power supply, the machine according to claim 3.

5. The control model includes a plurality of driver units arranged to determine priorities of a set of available actions, the plurality of driver units including a basic driver and a series of auxiliary drivers following it, the machine according to claim 4.

6. The machine according to claim 5, wherein the basic driver is configured to maintain a positive energy state of the machine.

7. The machine according to claim 4 or 5, wherein the first control element is configured to suspend one or more auxiliary drivers during an initial operation period.

8. The machine according to any one of claims 4 to 7, wherein the first control element includes an adaptive learning module configured to update the control model.

9. The machine according to any one of the preceding claims, wherein the common mode regulator is driven by an internal reference signal indicating a positive energy state of the machine.

10. The common mode bias maintained by the first control element is configured to provide positive internal power dissipation to the common mode output of the mutual interface, and the common mode regulator is configured to adjust the internal power dissipation to control the internal temperature of the machine, the machine according to any one of the preceding claims.

11. The first set of mutual information channels and the second set of mutual information channels each include a plurality of mutual channel pairs that establish an actual energy basis in the second control element, the machine according to any one of the preceding claims.

12. The non-reciprocal interface includes a non-reciprocal output, and the first set of mutual information channels further includes an output channel that transmits information to the non-reciprocal output and an emulated input channel that forms a mutual pair with the output channel, the machine according to any one of the preceding claims.

13. The first set and the second set of mutual information channels each include a plurality of non-reciprocal channel pairs, each channel pair consisting of a non-reciprocal channel provided from either the first control element or the physical interface and a corresponding emulated channel, the machine according to any one of the preceding claims.

14. One or more of the non-reciprocal channel pairs terminate in front of the first control element or the physical interface, thereby forming stub channels, the machine according to claim 13.

15. The first prediction model is configured to generate an output including signals in all of the first set of mutual information channels, and the second prediction model is configured to generate an output including signals in all of the second set of mutual information channels, the machine according to any one of the preceding claims.

16. The first prediction model and the second prediction model operate towards a convergence state in which the energy flux between the first set of mutual information channels and the second set of mutual information channels is minimized, and the second control element is configured to stepwise control the path to the convergence state, the machine according to any one of the preceding claims.

17. In a divergent state, the second control element is configured to identify an observable rendering of the energy flux between the first set of mutual information channels and the second set of mutual information channels at the physical interface and control the path to the convergence state based on the identified rendering, the machine according to claim 16.

18. The second control element is configured to adjust the path to the convergence state, the machine according to claim 17.

19. Composed in a distributed manner across a plurality of physical sub-components, the machine according to any one of the preceding claims.

20. The machine according to any one of the preceding claims, wherein the first control element is configured to generate a non-linear response to one or more of the second set of mutual information channels.

21. The machine according to any one of the preceding claims, wherein the second control element is configured to introduce perturbations into the first prediction model and / or the second prediction model.

22. A method of operating an autonomous machine according to any one of the preceding claims, comprising: determining an error energy flux between the first set of mutual information channels and the second set of mutual information channels based on observable characteristics of the physical interface; and adapting one or both of the first prediction model and the second prediction model based on the determined error energy flux.

23. The method according to claim 22, wherein the first prediction model and the second prediction model each include a hierarchical model, and adapting one or both of the first prediction model and the second prediction model includes selecting a layer to add or modify in each hierarchical model to reduce the error energy flux.

24. A computer program product comprising computer-readable instructions stored on a non-transitory carrier, the computer-readable instructions being executable by a computer to perform the method according to claim 22 or 23.

Images