DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 04/22/2024. The submission follows the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections – 35 USC § 103
The following is a quotation of 35 U.S.C. 103, which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
Claims 1-2, 10-11, and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Johansson et. al. “The Truth is in There - Rule Extraction from Opaque Models Using Genetic Programming (year:2004)” (“Johansson”) in view of Carven et. al. “Extracting Tree-Structured Representations of Trained Networks (1994)” (“Carven”) and in view of Koza et. al “Genetic programming as a means for programming computers by natural selection (1994)” (“Koza”) and in further view of Hodjat et. al., pre-grant num-US 20170293849 A1 (“Hodjat”)
As to claim 1
Johansson teaches “memory storing program instructions; a processor executing instructions stored in the memory; and a rule-set model generation engine executed by the processor and configured to:” (Johansson, “[pg3] the study is a comparative one where the results from
G-REX are compared both to the original results (from the opaque model) and to the results from standard techniques The standard techniques are the default selections for the respective problem category in the data-mining tool Clementine². For classification tasks this is (boosted) decision trees using the C5.0³ algorithm. Additionally, footnote 2 details that the software is spss and footnote 3 details the use of C 5.0 computer algorithm to create boosted decision trees.
Examiner notes: Because Clementine and C5.03 algorithm are software that require a computer for execution it is inherent that a computer contains a processor and a memory and a rule -set generation engine within.
2. receive a set of inputs from an input unit, wherein the set of inputs comprises one or more pre-generated deep learning models. (Johansson, pg-3, pg-4 “[pg-4], Using these functions and terminal sets the feasible expressions are exactly the same for G-REX and C&RTrees. G-REX uses the results of the trained [receive a set of inputs from an input unit,] ANN [one or more pre-generated deep learning models] as fitness cases, i.e. the fitness is based on fidelity. In addition, a penalty term is applied to longer representations, thus enforcing more compact trees. [pg-3] “In addition G-REX will use not only ANNs to extract from, but also another opaque model, i.e. boosted decision trees.”).
Johansson does not explicitly teach
evaluate the set of inputs by querying the set of inputs with one or more pre-defined querying datasets
generate an output comprising one or more outcomes of the evaluation
map the output with each of the pre-defined querying datasets used for querying the deep learning model, wherein a new dataset is generated based on the mapping,
randomly generate a population of initial rule-set models based on a set of hyper parameters, wherein the hyper parameters relate to configuration parameters used for generating population of initial ruleset models.
carry out an evolution process on the generated rule-set models for evolving the rule-set models by using the generated new datasets.
and execute the evolved rule-set model to solve one or more real-world problems.
Craven teaches “evaluate the set of inputs by querying the set of inputs with one or more pre-defined querying datasets.” (Craven, [pg-1] We present a novel algorithm, TREPAN, for extracting comprehensible symbolic representations from trained neural networks. TREPAN queries a given network [evaluate the set of inputs by querying the set of inputs] to induce a decision tree that describes the concept represented by the network. We evaluate our algorithm using several real-world problem domains and present results that demonstrate that TREPAN is able to produce decision trees that are accurate and comprehensible and maintain a high level of fidelity to the networks from which they were extracted. […] [pg-25] As shown in Table 1, the oracle is used for three different purposes: (i) to determine the class labels for the network's training examples; [with one or more pre-defined querying datasets]
2. generate an output comprising one or more outcomes of the evaluation. (Carven, “pg-26
PNG
media_image1.png
156
714
media_image1.png
Greyscale
Examiner notes: one or more outcomes of the evaluation is interpreted as Oracle(E) for each queried input.”).
3. map the output with each of the pre-defined querying datasets used for querying the deep learning model, wherein a new dataset is generated based on the mapping, (Carven, “pg-26
PNG
media_image2.png
482
726
media_image2.png
Greyscale
Examiner notes: pre-defined querying datasets used for querying the deep learning model, are interpreted as training examples and is mapped to each output i.e. ORACLE(E) and new dataset is generated on mapping is interpreted as labeled dataset that the algorithm builds its tree from.”).
Carven and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Carven it would have been obvious for a person of ordinary skill in the art to apply the teachings of Carven to Johansson before the effective filing date of the claimed invention in order to improve interpretability, i.e. human readable decision, and the accuracy of rule set model, while increasing scalability in real world application. (Carven, “[abs] TREPAN, for extracting comprehensible, symbolic representations from trained neural networks. Our algo-rithm uses queries to induce a decision tree that approximates the concept represented by a given network. Our experiments demonstrate that TREPAN is able to produce decision trees that maintain a high level of fidelity to their respective networks while being com-prehensible and accurate. Unlike previous work in this area, our algorithm is general in its applicability and scales well to large networks and problems with high-dimensional input spaces.”).
Johansson, in view of Carven does not teach:
randomly generate a population of initial rule-set models based on a set of hyper parameters, wherein the hyper parameters relate to configuration parameters used for generating population of initial ruleset models
carry out an evolution process on the generated rule-set models for evolving the rule-set models by using the generated new datasets.
and execute the evolved rule-set model to solve one or more real-world problems.
Koza teaches “randomly generate a population of initial rule-set models based on a set of hyper parameters, wherein the hyper parameters relate to configuration parameters used for generating population of initial ruleset models”.(Koza, pg-91, pg-93, pg-94, “[pg-91] In summary, genetic programming breeds computer programs to solve problems by executing the following three steps: 1. Generate an initial population of random computer programs composed of the primitive functions and terminals of the problem. [pg-93] The fourth major step in using genetic programming is selecting the values of certain parameters. The two major parameters that are used to control the process are the population size M and the maximum number of generations Ngen to be run. Ngen is 51 throughout this article. our choice of 4000 as the population size for this problem. [randomly generate a population; hyper parameters relate to configuration parameters used for generating population]. As it happens, a total of 23 individuals out of the 4000 in this initial random population tied with the highest score of 1280 matches on generation 0. One of these 23 high-scoring individuals was the S-expression (IF A0 D1 D2).”) [of initial rule-set models]
Koza and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Koza it would have been obvious for a person of ordinary skill in the art to apply the teachings of Koza to Johansson before the effective filing date of the claimed invention in order to allow a rule set to adapt the new data distributions. (Koza, “[pg-88] As will be seen, this algorithm will produce populations of computer programs which, over many generations, tend to exhibit increasing average fitness in dealing with their environment. In addition, these populations of computer programs can rapidly and effectively adapt to changes in the environment.”).
Johansson in view of Carven and in view of Koza does not teach:
carry out an evolution process on the generated rule-set models for evolving the rule-set models by using the generated new datasets.
and execute the evolved rule-set model to solve one or more real-world problems.
Hodjat teaches “carry out an evolution process on the generated rule-set models for evolving the rule-set models by using the generated new datasets.” (Hodjat, paragraph [0090] After the candidate individual pool 116 [the generated rule-set models] has been updated, a procreation module 608 evolves. [carry out an evolution process on] a random subset of them. Only individuals in the candidate individual pool with high fitness scores are permitted to procreate. Any conventional or future-developed technique can be used for procreation. In an embodiment, conditions, outputs, or rules from parent individuals are combined in various ways to form child individuals, [evolving the rule-set models by using the new datasets generated.] and then, occasionally, they are mutated.
2. and execute the evolved rule-set model to solve one or more real-world problems. (Hodjat paragaraph “[0045] When one or more rulesets process this production input data, the ruleset as a whole outputs a probability [ and execute the evolved rule-set model] 126 that a blood-pressure related event will occur in the near future. In a financial asset trading environment, for example, the production data sequence 130 may be a stream of real time stock prices and the output probability 126 may be the probability that a stock price will soon decrease. [to solve one or more real-world problems]”).
Hodjat and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Hodjat it would have been obvious for a person of ordinary skill in the art to apply the teachings of Hodjat to Johansson before the effective filing date of the claimed invention in order to extract the meaning data that is better understood by humans, in wide range of applications to solve real practical problems. (Hodjat, “paragraph [0004] The invention relates generally to data mining, and more particularly, to the use of genetic algorithms to extract useful rules or relationships from a data set for use in controlling systems.”[0012] However, many different kinds of applications may benefit from this approach, such as other medical diagnoses or interventions, predicting/anticipating weather conditions, or predicting/anticipating an impending occurrence of a natural disaster. For example, a computer executed ruleset may be used to automatically provide input to a medical professional regarding whether treatment is needed given a set of symptoms and/or measured vital signs. Rather than having the ruleset prescribe a treatment directly, the ruleset provides a probability that a particular event will occur or that a particular state may be attributed to an entity from which the data was observed/collected.”)
As to Claim 11
Johansson teaches “A method for adding explainability to deep learning models by using rule-set evolution, the method is implemented by a processor executing program instructions stored in a memory, the method comprises:” (Johansson, abs, pg-3 “[abs] in this study we show how the gap between accuracy and other aspects can be bridged by using a rule extraction method (termed G-REX) based [ based on rule-set evolution, the system comprising] on genetic programming. [pg3] the study is a comparative one where the results from G-REX are compared both to the original results (from the opaque model) and to the results from standard techniques The standard techniques are the default selections for the respective problem category in the data-mining tool Clementine². For classification tasks this is (boosted) decision trees using the C5.0³ algorithm. Additionally, footnote 2 details that the software is spss and footnote 3 details the use of C 5.0 computer algorithm to create boosted decision trees.
Examiner notes: Because Clementine and C5.03 algorithm are software that requires a computer for execution it is inherent that a computer contains a processor and a memory within executing program instructions.
And for all the other limitations of Claim 11, it is rejected on the same basis as Claim 1. As the Claim are Analogous.
As to claim 20:
Johansson teaches A computer program product comprising: a non-transitory computer-readable medium having computer program code stored thereon, the computer-readable program code comprising instructions that, when executed by a processor, causes the processor to” :” (Johansson, [pg3] The study is a comparative one where the results from G-REX are compared both to the original results (from the opaque model) and to the results from standard techniques The standard techniques are the default selections for the respective problem category in the data-mining tool Clementine². For classification tasks this is (boosted) decision trees using the C5.0³ algorithm. Additionally, footnote 2 details that the software is spss and footnote 3 details the use of C 5.0 computer algorithm to create boosted decision trees.
Examiner notes: Because Clementine and C5.03 algorithm are software that require a computer for execution it is inherent that a computer contains a processor and a memory (having computer code) within, executing program instructions.
And for all the other limitations of Claim 20, it is rejected on the same basis as Claim 1. As the Claim are Analogous.
As to claim 2
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat of teaches the method of claim 1.
Craven further teaches “wherein the one or more pre-defined querying dataset types comprises training datasets, validation datasets, testing datasets which are used for generating the deep learning model and synthetically generated datasets.” (Craven, pg-1, pg-28 “ [pg-1] We present a novel algorithm, TREPAN, for extracting comprehensible symbolic representations from trained neural networks. TREPAN queries a given network [wherein the one or more pre-defined querying dataset types comprises] to induce a decision tree that describes the concept represented by the network. […] The neural networks we use in our experiments have a single layer of hidden units. The number of hidden units used for each network (0, 5, 10, 20 or 40) is chosen using cross validation on the network's training set, and we use a validation set [validation datasets,] to decide when to stop training networks. […] We measure accuracy and fidelity on the examples in the test sets. [testing datasets] Queries to the oracle can specify constraints, however, do not have to be complete instances, but instead on the values that the features can take. In the latter case, the oracle generates a complete instance by randomly selecting values for each feature, while ensuring that the constraints are satisfied. In order to generate these random values- [and synthetically generated datasets], TREPAN uses the training data to model each feature's marginal distribution. [training datasets]
Johansson, Carven, Koza and Hodjat are combinable for the same rationale as set forth above with respect to Claim 1
As to claim 10 and Analogous claim 19
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat teaches the system of Claim 1.
Johansson further teaches “and wherein the rule-set model duplicates the behavior of the deep learning models trained with a supervised dataset in the prediction or classification problems” (Johansson pg-3, pg-6 “[pg-3] The overall purpose of this study is to evaluate G-REX on new tasks and using new representation languages. More specifically G-REX will be extended to handle: • Regression problems producing regression trees. • Classification problems producing fuzzy rules. [or classification problems] In addition, G-REX will use not only ANNs to extract from, but also another opaque model, i.e. boosted decision trees. [pg-4] G-REX extracting fuzzy rules from ANNs. In this experiment the extracted rule is a fuzzy rule. Each input variable has been manually fuzzified and has two possible fuzzy values, labeled Low and High. [with a supervised dataset] […] To produce a prediction the output from the fuzzy rule is compared to a threshold value, which also evolved for each candidate rule. [pg-6] Fidelity. Although this is not the main purpose of the G-REX algorithm the study shows that the extracted representations [and wherein the rule-set model duplicates the behavior] have very similar performance to the ANNs [of the deep learning models,] both on training and test sets.
2. and wherein the rule-set models duplicate the performance of the deep learning models (Johansson, pg-6 [pg-6] Fidelity. Although this is not the main purpose of the G-REX algorithm the study shows that the extracted representations [and wherein the rule-set model] have very similar performance [duplicates the performance] to the ANNs [of the deep learning models,] both on training and test sets.
3. and wherein similar prescriptions occur when evaluation is done directly and through a surrogate model. (Johansson “[pg-6] Although this is not the main purpose of the G-REX algorithm the study shows that the extracted representations have very similar performance [and wherein similar prescriptions occur when evaluation is done directly] to the ANNs, [ surrogate model]. both on training and test sets.
Hodjat further teaches “wherein the evolved rule-set model is applied to solve the real-world problem comprising prediction or classification problems and prescription or action determination problems” (Hodjat [0045] When one or more rulesets process this production input data, the ruleset as a whole outputs a probability 126 [wherein the evolved rule-set model] that a blood-pressure related event will occur in the near future. In a financial asset trading environment [is applied], for example, the production data sequence 130 may be a stream of real time stock prices and the output probability 126 may be the probability that a stock price will soon decrease. [to solve the real-world problem comprising prediction or classification problems and prescription or action determination problems]
2. evolved to optimize the outcomes of prescriptions in the prescription or action determination problems (Hodjat “[0070] The decision/action system 128 is a system that uses the probability output 126 from the rulesets together with predetermined threshold values to decide what if any action to take. [or action determination problems]. In an embodiment, the decision/action system 128 may output a recommendation for a human to perform an action. [...] for example 50%, that a patient's blood pressure will exceed the normal range in the near future, the decision/action system 128 may alert a nurse or a doctor. In an embodiment, the decision/action system 128 may also recommend an action. For example, if the ruleset predicts a probability that a patient's blood pressure will exceed the normal range in the near future with a probability greater than the threshold probability of 50%, the decision/action system 128 may recommend administering a blood pressure lowering drug. [evolved to optimize the outcomes of prescriptions in the prescription].
Johansson, Carven, Koza and Hodjat are combinable for the same rationale as set forth above with respect to Claim 1
Claims 3-7, and 12-16 are rejected under 35 U.S.C. 103 as being unpatentable over (“Johansson”) in view (“Carven”) and in view of” (“Koza”) and in further view of (“Hodjat”) and in view of Babak et .al, pre-grant num- US 20180113977 A1 (“Babak”).
As to claim 3 and Analogous claim 12
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat of teaches the system of claim 1.
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat does not teach:
wherein the rule-set generation engine comprises a rule-set model evolution unit executed by the processor and configured to carry out the evolution process
by reproducing the rule-set models based on selection of parents from the generated rule-set models, and wherein the selection of parents is carried out by applying a genetic evolution technique
Babak teaches wherein the rule-set generation engine comprises a rule-set model evolution unit executed by the processor and configured to carry out the evolution process. (Babak, “paragraph [0060] The evolutionary engine creates, [wherein the rule-set generation engine comprises] tests, and harvests the best individuals to be used in production system 112. The host memory 426 contains, among other things, computer instructions which, when executed by the processor subsystem 414, cause the computer system to operate or perform functions as described herein. [a rule-set model evolution unit executed by the processor]”).
2. by reproducing the rule-set models based on selection of parents from the generated rule-set models, and wherein the selection of parents is carried out by applying a genetic evolution technique. (Babak “paragraph [0006] Individuals with the best fitness estimate are then used to create the next generation. Through procreation, rules of parent individuals [rule-set models] are mixed, and sometimes mutated (i.e., a random change is made in a rule) to create a new rule set. This new rule set is then assigned to a child individual that will be a member of the new generation. [0007] When a candidate is first created, it may do a bad job at solving the problem and hence have low fitness. But the genetic algorithm evolves the pool incrementally by discarding the least fit individuals, using the most fit individuals as parents in the procreation step to generate new individual by crossover and/or mutation [by reproducing based on selection of parents from the generated rule-set models] and repeating the evaluation of the new pool of candidate individual. The expectation is that after a large number of generations of this evolutionary process, [configured to carry out the evolution process] the fittest individuals then in the pool will embody the optimal solutions to the target problem
Babak and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Babak it would have been obvious for a person of ordinary skill in the art to apply the teachings of Babak to Johansson in view of Carven, Koza and Hodjat before the effective filing date of the claimed invention in order to improve performance, and adaptability in real-world applications while maintaining high level of transparency. (Babak,” paragraph [0027] In the healthcare domain, an individual (genome) might propose a diagnosis based on patient prior treatment and current vital signs, and fitness may be measured by the accuracy of that diagnosis as represented in the training data [0028] In a financial assets trading environment, the individual can be used to detect patterns in real time data and assert trading signals to a trading desk.”) .
As to claim 4 and Analogous claim 13
Johansson, in view of Carven, in view of Koza and in further view of Hodjat and in view of Babak of teaches the system of claim 3.
Hodjat further teaches “wherein the rule-set model evolution unit is configured to generate offsprings by applying a crossover technique and a mutation technique on the selected parents.” (Hodjat, [0090] After the candidate individual pool 116 has been updated, a procreation module 608 evolves a random subset of them. Only individuals in the candidate individual pool with high fitness scores are permitted to procreate [technique on the selected parents.] Any conventional or future-developed technique can be used for procreation. In an embodiment, conditions, outputs, or rules from parent individuals are combined in various ways to form child individuals, and then, occasionally, they are mutated. [a mutation] The combination process for example may include crossover [crossover technique] - i.e., exchanging conditions, outputs, or entire rules between parent individuals to form child individuals. [wherein the rule-set model evolution unit is configured to generate offsprings]
Johansson, Carven, Koza, Hodjat and Babak are combinable for the same rationale as set forth above with respect to Claim 3.
As to claim 5 and Analogous claim 14
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat and in view of Babak teaches the system of claim 4.
Koza further teaches “wherein a first crossover technique is applied by selecting a random crossover index less than the number of rule-set models in one parent individual and in the offspring rule-set model,” Koza [“pg-91”] (ii) Crossover: Create two new offspring programs for the new population by genetically recombining randomly chosen parts of two existing programs. [...] [pg-91]
PNG
media_image3.png
674
488
media_image3.png
Greyscale
with respect to fig. 4; as 4 details when i[ index less than] does not equal M[the number of rule-set models] the crossover is chosen probabilistically i.e. PC [wherein a first crossover technique is applied by selecting a random crossover]and chooses two individuals [in one parent individual], i is incremented by 1, crossover is performed and two offspring is inserted into the new population [and in the offspring rule-set model]
2. and replacing remainder rule-set models in that individual with rules past the crossover index from the other parent, (Koza, [pg-107] Moreover, since a constrained syntactic structure is involved, we must perform structure-preserving crossover so as to ensure the syntactic validity of all offspring as the run proceeds from generation to generation. Structure-preserving crossover is implemented by first allowing the selection of the crossover point in the first parent to be any point from the body of ADFO (type 9), ADFI (type 10), [and replacing remainder rule-set models in that individual] or the result-producing branch (type 11). However, once the crossover point in the first parent has been selected, the crossover point of the second parent must be of the same type (i.e. types 9, 10, or 11). . [with rules past the crossover index from the other parent] This restriction on the selection of the crossover point of the second parent assures syntactic validity of the offspring.
3. and wherein a second crossover technique is applied by carrying out a logical multiplication of one parent rule-set model into a second parent for producing offspring with longer rules than the parents. (koza [pg-90] The crossover operation [and wherein a second crossover technique] creates new offspring [for producing offspring] by exchanging subtrees (i.e. sublists, subroutines, sub procedures) between the two parents. Assume that the points of both trees are numbered in a in a depth-first way starting at the left. Suppose that the point number 2 (out of 7 points of the first parent) is randomly selected as the crossover point for the first parent and that the point number 5 (out of 9 points of the second parent) is randomly selected as the crossover point of the second parent. The crossover points in the trees above are therefore the * in the first parent and + in the second parent. The two crossover fragments are the two sub-treesshown in Fig. 2. These two crossover fragments correspond to the underlined subprograms (sublists) in the two parental computer programs. The two offspring resulting from crossovers are
PNG
media_image4.png
170
698
media_image4.png
Greyscale
[by carrying out a logical multiplication of one parent rule-set model into a second parent with longer rules than the parents]
Johansson, Carven, Koza, Hodjat and Babak are combinable for the same rationale as set forth above with respect to Claim 3.
As to claim 6 and Analogous claim 15
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat and in view of Babak of teaches the system of claim 4.
Hodjat further teaches “wherein the mutation technique is applied by a rule-set model evolution unit by carrying out a single random change in an element of the rule-set model”. (Hodjat, abs, paragraph,008 “[abs]. Described is a ruleset which predicts the probability of a particular outcome. [a rule-set model evolution unit] Roughly described, an individual identifies a ruleset, where each of the rules has a plurality of conditions and also indicates a rule-level probability of a predetermined classification. [0008] Through procreation, rules of parent individuals are mixed, and sometimes mutated (i.e., a [wherein the mutation technique is applied by a rule-set model evolution unit] random change is made in a rule) to [carrying out a single random change in an element of the rule-set model,]. create a new ruleset [0098] The rule-level probability 810 indicates the probability of class membership when the conditions of this rule are satisfied. A rule-level probability 810 may be selected by user input, chosen at random, etc. In an embodiment, the rule-level probability is one of set of enumerated probability values, such as 0.1, 0.2, 0.3 . . . 0.9[0099] During procreation, any of the conditions 308 or the rule-level probability 810 may be altered, or even entire rules may be replaced. The individual's fitness estimates 804 are determined by the candidate testing module 604 in training system 112, accrued over all the trials.”).
2 and wherein the mutation technique is applied at the condition level by changing an element of the condition, or at the rule level by replacing, removing, or adding a condition to the rule-set model, or changing the rule-set model’s action. (Hodjat, abs, paragraph,008 “[abs]. Described is a ruleset which predicts the probability of a particular outcome Roughly described, an individual identifies a ruleset, where each of the rules has a plurality of conditions and also indicates a rule-level probability of a predetermined classification. [0008] Through procreation, rules of parent individuals are mixed, and sometimes mutated (i.e., a [wherein the mutation technique is applied] random change is made in a rule) to create a new ruleset; [or adding a condition to the rule-set model] [0098] The rule-level probability 810 indicates the probability of class membership when the conditions of this rule are satisfied. A rule-level probability 810 may be selected by user input, chosen at random, etc. In an embodiment, the rule-level probability is one of set of enumerated probability values, such as 0.1, 0.2, 0.3 . . . 0.9[0099] During procreation, any of the conditions 308 or the rule-level probability 810 may be altered, or even entire rules may be replaced. , [ at the condition level by changing an element of the condition, or at the rule level by replacing, removing, or changing the rule-set model’s action]The individual's fitness estimates 804 are determined by the candidate testing module 604 in training system 112, accrued over all the trials.”).
Johansson, Carven, Koza, Hodjat and Babak are combinable for the same rationale as set forth above with respect to Claim 3.
As to claim 7 and Analogous claim 16
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat and in view of Babak of teaches the system of claim 6.
Hodjat further teaches “wherein the mutation technique is applied at the rule-set model level by removing an entire rule set from the parent, or changing the default ruleset, or changing the rule-set order.” (Hodjat, paragraph, 0055, 0099, “[0055], A ruleset 300 is composed of one or more rules 306. Each rule 306 contains one or more conditions 308 and a rule-level probability (RLP) 310. A rule-level probability 810 is assigned to a rule at the time when the individual is created and is subject to evolution. The rule-level probability 810 indicates that the probability of class membership when the conditions of this rule are satisfied. A rule-level probability 810 may be selected by user input, chosen at random, etc. [0099] During procreation, [wherein the mutation technique is applied] any of the conditions 308 or the rule-level probability 810 may be altered, [or changing the default ruleset, or changing the rule-set order.], or even entire rules may be replaced. [is applied at the rule-set model level by removing an entire rule set from the parent] The individual's fitness estimates 804 is determined by the candidate testing module 604 in the training system 112, accrued over the all the trials.
Johansson, Carven, Koza, Hodjat and Babak are combinable for the same rationale as set forth above with respect to Claim 3.
Claim 8 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over (“Johansson”) in view (“Carven”) and in view of” (“Koza”) and in further view of (“Hodjat”) and in view of (“Babak”) and in view of Liang, et.al pre grant num- US 20190180186 A1, (“Liang”)
As to claim 8 and Analogous claim 17
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat and in view of Babak of teaches the system of claim 4.
Babak further teaches “and wherein a time-applied counter technique is associated with each ruleset for keeping track of a number of times” (Babak [0053] To determine activity percentages for one or more individuals, additional data is recorded at testing time. An individual activity level counter [and wherein a time-applied counter technique] may be incremented for each testing sample during which the individual produced an output (i.e., one or more rules in the ruleset fired). [is associated with each ruleset] For example, the testing process may record data for each sample including [for keeping track of a number of times] a sample identifier, an individual identifier, one or more rule identifiers of rules that fired for this sample, and for each rule that fired, the corresponding output from the rule.
The ruleset was evaluated as ‘true’ during evaluation (Babak [0020] Each individual includes an identifier, an indication of the amount of testing it has undergone, and the individual's current fitness estimate based on its performance on the testing data. An individual also includes one or more “rules”, each [the ruleset was evaluated as] of which contains one or more conditions and an output to be asserted if all the conditions in a given input data sample are true. [‘true’ during evaluation]
[and wherein another bloat-control technique utilizes the time-applied counter technique] to filter inactive individual rulesets from participating in crossover. (Babak, “paragraph [0052] Activity thresholds can also be applied to the activity of each rule within an individual. Failure to meet the minimum activity requirement may cause the candidate to be removed from the candidate pool. In one embodiment, when the minimum activity level is applied to activity of particular rules within an individual, just the inactive rules may be removed from the individual [filter inactive individual rulesets] and the rest of the individual's genetic material retained to create a new individual that is introduced as a new candidate with no experience (not tested on samples yet) back into the candidate pool to be tested again. [0007] But the algorithm evolves the pool incrementally by discarding the least fit individuals, using the most fit individuals as parents in the procreation step to generate new individuals by Crossover [from participating in crossover] and/or mutation and repeating the evaluation of the new pool of candidate individuals. The expectation is that after a large number of generations of this evolutionary process, the fittest individuals then in the pool will embody the optimal solutions to the target problem.”).
Examiner notes: [and wherein another bloat-control technique utilizes the time-applied counter technique]” (Babak paragraph [0053] discloses times-applied counter technique and reducing the bloat is disclosed by liang paragraph [0089].
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat and in view of Babak does not teach.
wherein the mutation technique generates a bloat by making the offspring smaller or larger than the original parent
and wherein in order to reduce the bloat all conditions recognized as falsehoods are removed from the offsprings
Liang teaches “wherein the mutation technique generates a bloat by making the offspring smaller or larger than the original parent” (Liang, paragraph, [103] In some implementations, the procreation module, in forming new genomes, forms certain new genomes by mutation [mutation technique] which adds a new submodule to a pre-existing module and/or supermodule. [generates a bloat by making the offspring [smaller] or larger than the original parent]. [0105] In some implementations, the procreation module, in forming new genomes, forms certain new genomes by mutation which deletes a pre-existing module and/or supermodule from a pre-existing genome. [making the offspring smaller or larger than the original parent].
Examiner notes: Under BRI, generating a bloat, making offspring bigger than the parent is interpreted as adding the new submodule during mutation makes the resulting offspring bigger than the parent. Additionally, deleting the new submodule during mutation makes the resulting offspring smaller than the parent, i.e bloat. (Specification-paragraph- [0024])
and wherein in order to reduce the bloat all conditions recognized as falsehoods are removed from the offsprings (Liang [0089] In other implementations, the competition module discards genomes [are removed from the offsprings] that do not meet the minimum baseline [recognized as falsehoods] genome fitness or whose “genome fitness” relatively lags the “genome fitness” of similarly tested genomes. Candidate genome pool database 902 is updated with the revised contents.
Examiner notes: Under BRI, in light of specification [Paragraph, [0024], reducing bloat is interpreted as genomes being discarded and recognizing the falsehoods is interpreted as genomes that do not meet the minimum baseline.”).
Liang and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Liang it would have been obvious for a person of ordinary skill in the art to apply the teachings of Liang to Johansson before the effective filing date of the claimed invention in order to select best fitness individual by removing falsehoods to improve convergence speed and overall performance of the Evolutionary process. Additionally, one skilled in the art would be motivated to protect integrity, accuracy and trustworthiness of evolved rule sets ensuring its application in various real-world domains. (Liang [0082] The production system 934 applies these genomes to production data, and produces outputs, which may be action signals or recommendations. In the financial asset trading environment, for example, the production data may be a stream of real time stock prices and the outputs of the production system 934 may be the trading signals or instructions that one or more of the genomes in the production genome pool 932 outputs in response to the production data.”).
Claim 9 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over (“Johansson”) in view (“Carven”) and in view of” (“Koza”) and in further view of (“Hodjat”) and in view of Bengio et.al, “Curriculum Learning (year-2009)”, (“Bengio”)
As to claim 9 and Analogous claim 18
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat teaches the system of Claim 1.
Johansson, in view of Carven, and in view of Koza and in further view of Hodjat does not teach:
wherein an expression vocabulary associated with the evolved rule-set models is expanded during evolution by implementing curricular learning.
Bengio teaches “wherein an expression vocabulary associated with the evolved rule-set models is expanded during evolution by implementing curricular learning.” (Bengio, pg-46,47 “[Pg-46] We chose the training set S as all possible windows of text of size n = 5 from Wikipedia (http://en.wikipedia.org), obtaining 631 million windows processed as in Collobert and Weston (2008). We chose as a curriculum strategy to grow [during evolution by implementing curricular learning.] the vocabulary size: [wherein an expression vocabulary associated with the evolved rule-set models is expanded] the first pass over Wikipedia was performed using the 5,000 most frequent words in the vocabulary, which was then increased by 5,000 words at each subsequent pass-through Wikipedia. At each pass, any window of text containing a word not in the considered vocabulary was discarded.
Bengio and Johansson are related to the same field of endeavor (Explainable AI). In view of the teachings of Bengio it would have been obvious for a person of ordinary skill in the art to apply the teachings of Bengio to Johansson before the effective filing date of the claimed invention to implement curricular learning so that the evolved rulesets model is more accurate, interpretable and faster. (Bengio, “[pg-46] To eliminate the explanation that better results are obtained with the curriculum because of seeing more examples, we trained a no-curriculum model with the union of the BasicShapes and GeomShapes training sets, with a final test error still significantly worse than with the curriculum (with errors similar to “switch epoch”=16). We also verified that training only with BasicShapes yielded poor results. We are interested here in training a language model, predicting the best word which can follow a given context of words in a correct English sentence. [pg-47] curriculum learning adds the notion of guiding the optimization process, either to converge faster, or more importantly, to guide the learner towards better local minima.”).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NIROJ KOIRALA whose telephone number is (571)270-0748. The examiner can normally be reached Monday -Friday 8am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, MICHAEL HUNTLEY can be reached on (303) 297-4307. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/N.K./Examiner, Art Unit 2129
/ADAM C STANDKE/Primary Examiner, Art Unit 2129
Prosecution Insights