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 .
Remarks
This Office Action is responsive to Applicants' Amendment filed on 04/30/2026, in which claims 1, 3-6, 8-15, and 17-20 have been amended. No claims are cancelled and no new claims are added.
Claims 1-20 are currently pending.
Response to Arguments
With regards to the rejections of claims 1-20 under 35 U.S.C. 101 for being directed towards abstract ideas, Examiner respectfully disagrees with Applicant’s arguments that the claims as amended overcome the rejections. Applicant argues with respect to claim 1.
Applicant argues first on page 10 and 11 of the Remarks that claim 1 is eligible at least at Step 2A, Prong One by not reciting a mental process, as it recites the limitation training, using reinforcement learning and the MDP, an intra-option policy for each of the parameterized options, which Applicant asserts is not a mental process as it cannot be practically performed within a human mind. Applicant cites MPEP 2106.04(a)(2)(III)(A), which states “[c]laims do not recite a mental process when they do not contain limitations that can practically be performed in the human mind, for instance when the human mind is not equipped to perform the claim limitations”.
Examiner respectfully disagrees that claim 1 does not recite any mental processes. While Examiner agrees that a human mind cannot perform training of intra-option policies, MPEP 2106.04(a) recites:
“Examiners should determine whether a claim recites an abstract idea by (1) identifying the specific limitation(s) in the claim under examination that the examiner believes recites an abstract idea, and (2) determining whether the identified limitations(s) fall within at least one of the groupings of abstract ideas listed above…If the identified limitation(s) falls within at least one of the groupings of abstract ideas, it is reasonable to conclude that the claim recites an abstract idea in Step 2A Prong One”.
As identified in the 101 rejection of claim 1 below, claim 1 recites numerous limitations that fall under the mental process grouping of abstract ideas, such as generating a mapping function between MDP states of a MDP within the MDP distribution and planning states of the planning domain. Other limitations that are not identified as abstract ideas, such as the limitation on training intra-option policies, are further evaluated at Step 2A, Prong Two and Step 2B, where they can potentially, but not necessarily, result in the claim as a whole being 101 eligible.
Applicant further argues on pages 10-12 of the Remarks that claim 1 is eligible at least at Step 2A, Prong Two by integrating any recited abstract ideas into a practical application “of policy generalization across a plurality of MDPs”. Applicant again recites their limitation on training intra-option policies as an example of a limitation that integrates into a practical application. Applicant refers to MPEP 2106.04(d)(I), stating: “Section 2106.04(d)(I) of the MPEP also states that ‘an improvement to other technology or technical field’ is ‘indicative that an additional element (or combination of elements) may have integrated the exception into a practical application’”. Applicant refers to paragraphs [0040] and [0056] of their Specification as disclosing an improvement of “use of a generalized intra-option policy that is compatible with multiple MDPs”.
Although Examiner does not dispute that the invention provides the described improvement, Examiner notes that MPEP 2106.04(d).III. states “Because a judicial exception alone is not eligible subject matter, if there are no additional claim elements besides the judicial exception, or if the additional claim elements merely recite another judicial exception, that is insufficient to integrate the judicial exception into a practical application”, i.e. that an improvement provided solely by improving an abstract idea itself is ineligible. Examiner further notes that the limitations of claim 1 as amended that are not identified as reciting abstract ideas either amount to mere data gathering, or in the case of the training limitation, is mere instructions to apply training and reinforcement learning, neither of which integrate the recited judicial exceptions into a practical application, as described in the 101 rejections below.
With regards to the rejections of claims 3, 10, and 17 under 35 U.S.C 112(b), Applicant has amended the claims to remove the previously noted causes of indefinite scope, and thus the rejections are withdrawn.
With regards to the rejections of claims 1-3, 6-10, 13-17, and 20 under 35 U.S.C. 103 as being unpatentable over Lee et al. “AI Planning Annotation for Sample Efficient Reinforcement Learning”, in view of Konidaris and Doshi-Velez “Hidden Parameter Markov Decision Processes: An Emerging Paradigm for Modeling Families of Related Tasks”, further in view of Juba et al. “Safe Learning of Lifted Action Models”, Examiner finds Applicant’s arguments that the claims as amended overcome the rejections are persuasive, however the arguments are moot in view of a new grounds of rejection, necessitated by Applicant’s amendments to the claims, as presented below.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a No therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract ideas without significantly more.
Regarding claim 1,
Step 1 - “Is the claim to a process, machine, manufacture or composition of matter?”
Yes, the claim is directed towards a process.
Step 2A, Prong 1 - “Is the claim directed to a law of nature, a natural phenomenon (product of nature) or an abstract idea?”:
The limitation of generating a mapping function between MDP states of a MDP within the MDP distribution and planning states of the planning domain; recites an evaluation of a mapping function, which is a mental process, which is an abstract idea, regardless of if it’s performed on a generic computer or using a generic machine learning model.
The limitation of defining, using the mapping function, a parameterized option for each of the lifted action models, wherein the parameterized option corresponds to a plurality of actions defined by a lifted action model of the lifted action models, recites an evaluation of options and actions, which is a mental process, which is an abstract idea, regardless of if it’s performed on a generic computer or using a generic machine learning model.
The limitation of and wherein the parameterized option is based at least in part on one or more parameters, wherein the one or more parameters include lifted first-order predicates; recites an evaluation of parameters to define an option, which is a mental process, which is an abstract idea, regardless of if it’s performed on a generic computer or using a generic machine learning model.
Step 2A, Prong 2 - “Does the claim recite additional elements that integrate the judicial exception into a practical application?”:
The limitation of receiving a planning domain including lifted action models; recites the mere extra-solution activity of data gathering, which does not integrate the exception into a practical application, MPEP 2106.05(d) and 2106.05(g).
The limitation of receiving a Markov Decision Process (MDP) distribution; recites the mere extra-solution activity of data gathering, which does not integrate the exception into a practical application, MPEP 2106.05(d) and 2106.05(g).
The limitation of and training, using reinforcement learning and the MDP, an intra-option policy for each of the parameterized options, wherein the intra-option policy is compatible with an additional MDP recites mere instructions to apply reinforcement learning to train intra-option policies that are generalizable, which does not integrate the exception into a practical application, MPEP 2106.05(d) and 2106.05(f).
Step 2B - “Does the claim recite additional elements that amount to significantly more than the judicial exception?”:
The limitation of receiving a planning domain including lifted action models; recites receiving data over a network, which is well-understood, routine, and conventional, MPEP 2106.05(d).II., example (i) of WURC computer functions.
The limitation of receiving a Markov Decision Process (MDP) distribution; recites receiving data over a network, which is well-understood, routine, and conventional, MPEP 2106.05(d).II., example (i) of WURC computer functions.
The limitation of and training, using reinforcement learning and the MDP, an intra-option policy for each of the parameterized options, wherein the intra-option policy is compatible with an additional MDP recites mere instructions to apply reinforcement learning to train intra-option policies that are generalizable, which is not significantly more than any recited judicial exceptions, MPEP 2106.05(d) and 2106.05(f).
Therefore, claim 1 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 2,
Claim 2 adds the additional limitation to claim 1:
wherein a particular one of the parameterized options is defined as: an initiation set for the particular one of the parameterized options, one or more option parameters, a termination condition for the particular one of the parameterized options, and an intra-option policy for the particular one of the parameterized options recites further detail on the parameterized options that are defined with a mapping function, without changing that the evaluation of the mapping function amounts to the abstract idea of evaluation, regardless of if it’s performed on a generic computer.
Therefore, claim 2 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 3,
Claim 3 adds the additional limitation to claim 2:
initializing a replay buffer for the particular one of the parameterized options, recites mere instructions to apply a replay buffer for an option, which does not integrate the exceptions into a practical application, and is not significantly more than any recited judicial exceptions, MPEP 2106.05(d) and 2106.05(f).
the training for the particular one of the parameterized options includes storing, within the replay buffer and for a particular action from the intra-option policy for the particular one of the parameterized options, data including: an initial state, the particular action, a reward, a subsequent state, and one or more values associated with the one or more option parameters, recites the mere extra-solution activity of selecting a particular type of data to be manipulated, which does not integrate the exception into a practical application, MPEP 2106.05(d) and 2106.05(g), and which recites storing information in memory, which is well-understood, routine, and conventional, MPEP 2106.05(d).II., example (iv) of WURC computer functions.
and the intra-option policy for the particular one of the parameterized options is updated using the data recites mere instructions to apply a data to update a policy, which does not integrate the exceptions into a practical application, and is not significantly more than any recited judicial exceptions, MPEP 2106.05(d) and 2106.05(f).
Therefore, claim 3 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 4,
Claim 4 adds the additional limitations to claim 1:
the MDP distribution defines an environment including a plurality of MDPs meeting constraints, recites an evaluation of an environment, which is a mental process, which is an abstract idea, regardless of if it’s performed on a generic computer or using a generic machine learning model.
and the constraints including a predicate, action, object, type, and action model recites further detail on the constraints within an environment, without changing that defining an environment amounts to the abstract idea of evaluation, regardless of if it’s performed on a generic computer.
Therefore, claim 4 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 5,
Claim 5 adds the additional limitation to claim 4:
wherein the policy for performing the goal is configured to be used with the additional MDP, wherein the additional MDP meets the constraints recites further detail on the policy that is trained, without changing that generic training of a policy amounts to mere instructions to mere instructions to apply, which does not integrate the exceptions into a practical application, and is not significantly more than any recited judicial exceptions, MPEP 2106.05(d) and 2106.05(f).
Therefore, claim 5 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 6,
Claim 6 adds the additional limitation to claim 1:
wherein the mapping function is generated using planning annotated reinforcement learning (PaRL) recites mere instructions to apply a planning annotated reinforcement learning to generate a mapping, which does not integrate the exceptions into a practical application, and is not significantly more than any recited judicial exceptions, MPEP 2106.05(d) and 2106.05(f).
Therefore, claim 6 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claim 7,
Claim 7 adds the additional limitation to claim 1:
wherein the planning domain includes a plurality of options for performing the goal recites further detail on the planning domain that is received, without changing that receiving a planning domain amounts to the mere extra-solution activity of data gathering, which does not integrate the exception into a practical application, MPEP 2106.05(d) and 2106.05(g), and which recites receiving data over a network, which is well-understood, routine, and conventional, MPEP 2106.05(d).II., example (i) of WURC computer functions.
Therefore, claim 7 is found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claims 8-14,
Claims 8-14 recite a computer hardware system, which is a machine, for performing the function of the method of claims 1-7, respectively, with substantially the same limitations. Therefore the same analysis and rejection applied to claim 1-7 applies to claims 8-14.
Therefore, claims 8-14 are found to be ineligible subject matter under 35 U.S.C. 101.
Regarding claims 15-20,
Claims 15-20 recite a computer program product, comprising a computer readable storage medium, which is a manufacture, for performing the function of the method of claims 1-6, respectively, with substantially the same limitations. Therefore the same analysis and rejection applied to claim 1-6 applies to claims 15-20.
Therefore, claims 15-20 are found to be ineligible subject matter under 35 U.S.C. 101.
Prior Art
The following references are used for prior art claim rejections:
Lee et al. “AI Planning Annotation for Sample Efficient Reinforcement Learning”, hereinafter Lee
Konidaris and Doshi-Velez “Hidden Parameter Markov Decision Processes: An Emerging Paradigm for Modeling Families of Related Tasks”, hereinafter Konidaris
Juba et al. “Safe Learning of Lifted Action Models”, hereinafter Juba
Ng and Petrick “Generalised Task Planning with First-Order Function Approximation”, hereinafter Ng
Horcik and Fiser “Endomorphisms of Lifted Planning Problems”, hereinafter Horcik
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.
Claims 1-3, 6-10, 13-17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Lee in view of Konidaris, further in view of Juba, further in view of Ng.
Regarding claim 1,
Lee teaches A computer-implemented method for generating a policy for performing a goal ((Lee Pg. 2) “In this goal oriented environment, we are interested in the sparse reward task where the r(s, a) is a constant step cost when s ∉ G and zero in a goal state, and the objective is to learn a stationary optimal policy π* that maximizes the expected return, π*”) and including a plurality of intra-option policies, ((Lee Pg. 8) “We design a general method for injecting intrinsic rewards to RL agents from the abstract planning task by reformulating the underlying decomposed sub-MDPs with constraints visible to planning agents. Learning only the (intra-option) policies for these sub-MDPs is shown to work well in practice on various problems, significantly improving sample efficiency”) comprising:
receiving a planning domain [including lifted action models]; ((Lee Pg. 5)) “We also constructed the PaRL task E = <M, Π, L> by modeling a planning task based on the domain knowledge with an appropriate state mapping function L”, Lee does not teach lifted action models)
receiving a Markov Decision Process (MDP) [distribution]; ((Lee Pg. 2) “In reinforcement learning, we assume that an agent interacts with a goal oriented MDP M = <S, 𝒜, P, r, s0, G, γ> with a set of states S, a set of actions 𝒜, a state transition function P : S × 𝒜 × S → [0, 1], a reward function r : S × A → R, an initial state s0 ∈ S, a set of goal states G ⊂ S, and a discounting factor γ ∈ (0, 1) for the rewards”, Lee does not teach a distribution for the Markov Decision Processes)
generating a mapping function between MDP states of a MDP [within the MDP distribution] and planning states of the planning domain; ((Lee Pg. 3) “A PaRL task is a triple E := <M, Π, L>, where M := <S, A, P, r, s0, G, γ> is a goal-oriented MDP over RL states S, Π := <V, O, s’0, s∗> is a planning task over planning states S, and L : S → S’ is a surjective mapping from the states of the MDP S to planning task S’ satisfying s’0 = L(s0) and s∗ consistent with L(s) for all s ∈ G”, the PaRL task includes a mapping L between MDP S states and planning task S’, Lee does not teach an MDP distribution)
and training, using reinforcement learning and the MDP, an intra-option policy ((Lee Pg. 5) “RL agents also have the freedom to train the SMDP policy µ and intra-option policies πOo using either off-policy or on-policy algorithms”, (Lee Pg. 4) “solving a PaRL task can be divided into three steps:...Second, train each option policy with a frame constrained option MDP”) for each of the [parameterized] options, (Lee Pg. 5, Algorithm 1 shows training of policies for each option on lines 15 and 16)
PNG
media_image1.png
700
665
media_image1.png
Greyscale
Konidaris teaches the following further limitation that Lee does not completely teach:
receiving a Markov Decision Process (MDP) distribution; ((Konidaris Pg. 1) “HIP-MDPs model a distribution of tasks where the variation in the dynamics (or the reward function, though we do not discuss that case here) across the family of tasks can be captured by a set of hidden parameters, θ, encountered with probability Pθ”, (Konidaris Pg. 2) “HIP-MDPs have some useful properties which suggest that they merit study in their own right. One is that any specific task instance is an MDP, and can be solved independently as one”)
At the time of filing, one of ordinary skill in the art would have motivation to combine Lee and Konidaris by taking the method for creating mappings between Markov Decision Process (MDP) states and planning states to train intra-option policies, taught by Lee, and including the MDPs being from an MDP distribution, taught by Konidaris, as Konidaris teaches: (Konidaris Pg. 2) “However, HIP-MDPs have some useful properties which suggest that they merit study in their own right. One is that any specific task instance is an MDP, and can be solved independently as one…This allows us to model scenarios where the source tasks for transfer learning can be solved directly by reinforcement learning. Moreover, the hidden parameters θ are constant for each task…These properties allow us to gather large batches of transition data that correspond to fixed settings of θ, easing model learning. Finally, the latent parameters are a sufficient statistic for specifying an individual task…Thus, if we have a model of the HIP-MDP, we can expend great computational effort offline to solve various task instances, and then synthesize a parameterized policy (da Silva, Konidaris, and Barto 2012) to be deployed given the agent’s belief over θb at runtime”, that is, that learning policies for MDPs in an MDP distribution allows for policies to be learned faster, and for that learning to be transferred between MDPs within the distribution, both of which reduce the time and cost of learning. Such a combination would be obvious.
Juba teaches the following further limitations that neither Lee, nor Konidaris teach:
receiving a planning domain including lifted action models; ((Juba Pg. 3) “A classical planning domain is defined by a tuple <T, F, 𝒜, M> where T is a set of types, F is a set of lifted fluents, 𝒜 is a set of lifted actions, and M is an action model for 𝒜”)
defining, using the mapping function, a parameterized option for each of the lifted action models, wherein the parameterized option corresponds to a plurality of actions defined by a lifted action model of the lifted action models, ((Juba Pg. 2) “A lifted action A ∈ 𝒜 is a pair <name, params> where name is a symbol and params is a list of types, denoted name(A) and params(A), respectively, and arity(A, t) denotes the number of type-t parameters. The action model M for a set of actions A is a pair of functions preM and effM that map every action in 𝒜 to its preconditions and effects”, a set of actions corresponds to an option)
and wherein the parameterized option is based at least in part on one or more parameters, ((Juba Pg. 2) “To define the preconditions and effects of a lifted action, we first define the notion of a parameter-bound literal. A parameter binding of a lifted literal L and an action A is a function bL,A : params(L) → params(A) that maps every parameter of L to a parameter in A”, each action of the option being based on parameters corresponds to the option being based on one or more parameters)
At the time of filing, one of ordinary skill in the art would have motivation to combine Lee, Konidaris, and Juba by taking the method for creating mappings between Markov Decision Process (MDP) states and planning states to train intra-option policies, with the MDPs being from a distribution, jointly taught by Lee and Konidaris, and including lifted action models in the planning domain which are each mapped to parameterized options based on parameters, taught by Juba, as Juba teaches: (Juba Pgs. 1-2) “Stern and Juba (2017) proposed a sound algorithm for safe model-free planning,…However, their positive result is limited to grounded domain models, that is, domains that are not defined by lifted, i.e., parameterized, actions and fluents. The size of a grounded domain model can be arbitrarily larger than its corresponding lifted domain model. In particular, a single lifted action can yield a number of grounded actions that grow polynomially with the number of objects in the domain, with the number of parameters of the lifted action as its exponent. In addition, learning a grounded domain model limits the generalization possible between different groundings of the same lifted domain. For example, a grounded action model for a blocksworld domain with 8 blocks cannot be used to solve problems for a blocksworld domain with 9 blocks”, that is, that lifted action models that support the inclusion of parameters are significantly more efficient and applicable to a wider variety of problems than their alternative. Such a combination would be obvious.
Ng teaches the following further limitation that Juba does not explicitly teach and that neither Lee, nor Konidaris teach:
wherein the one or more parameters include lifted first-order predicates; ((Ng Pg. 1) “Relational reinforcement learning (RRL) learns in a first-order representation rather than a propositional representation…we propose a model-free RRL method…we use lifted state predicates, or first-order features, to approximate the Q-values of lifted actions rather than ground actions”, (Ng Pg. 11) “We describe the initialisation of a set of first-order features, Φˆa, for each lifted action ˆa as shown in Algorithm 4”, Algorithm 4, Pg. 12 shows that a set of actions has a set of first order features (synonymous to lifted predicates in Ng) initialized for each constituent action, a set of actions corresponds to an option)
PNG
media_image2.png
198
677
media_image2.png
Greyscale
At the time of filing, one of ordinary skill in the art would have motivation to combine Lee, Konidaris, Juba, and Ng by taking the method for creating mappings between Markov Decision Process (MDP) states and planning states to train intra-option policies, with the MDPs being from a distribution, and with lifted action models in the planning domain which are each mapped to parameterized options based on parameters, jointly taught by Lee, Konidaris, and Juba, and including lifted first-order predicates as parameters for the options, taught by Ng, as Ng teaches: (Ng Pg. 1) “Since state predicates are ground over objects, problems with different objects or numbers of objects have different sets of state predicates, or feature spaces, and generalisation is not possible. Instead, we use lifted state predicates, or first-order features, to approximate the Q-values of lifted actions rather than ground actions”, that is, that use of lifted state predicates increases the ability to generalize the results of the reinforcement learning. Such a combination would be obvious.
Regarding claim 2,
Lee, Konidaris, Juba, and Ng jointly teach The method of claim 1,
Lee further teaches:
wherein a particular one of the [parameterized] options is defined as: ((Lee Pg. 2) “A set of options 𝒪 formalizes the temporally extended actions that defines a semi-MDP (SMDP) over the original MDP M. A Markovian option O ∈ 𝒪 is a triple <IO, πO, βO>”, Juba but not Lee teaches parameters for options)
an initiation set for the particular one of the [parameterized] options, ((Lee Pg. 2) “A Markovian option O ∈ 𝒪 is a triple <IO, πO, βO>, where the IO is the initiation set in which O can begin”, Juba but not Lee teaches parameters for options)
a termination condition for the particular one of the [parameterized] options, ((Lee Pg. 2) “A Markovian option O ∈ 𝒪 is a triple <IO, πO, βO>, where…the βO is a termination set in which O terminates”, Lee does not teach parameters for options)
and an intra-option policy for the particular one of the [parameterized] options ((Lee Pg. 2) “A Markovian option O ∈ 𝒪 is a triple <IO, πO, βO>, where…the πO is a stationary option policy πO : S × A → [0, 1]”, Juba but not Lee teaches parameters for options)
Juba further teaches:
one or more option parameters, ((Juba Pg. 2) “A lifted action A ∈ 𝒜 is a pair <name, params> where name is a symbol and params is a list of types, denoted name(A) and params(A), respectively, and arity(A, t) denotes the number of type-t parameters. The action model M for a set of actions A is a pair of functions preM and effM that map every action in 𝒜 to its preconditions and effects”, a set of actions corresponds to an option)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Lee, Konidaris, Juba, and Ng for the parent claim of claim 2, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 3,
Lee, Konidaris, Juba, and Ng jointly teach The method of claim 2, further comprising:
Lee further teaches:
initializing a replay buffer for the particular one of the [parameterized] options, (Lee Pg. 5, Algorithm 1 instructs on line 1 to “Initialize trajectory buffer B”, and on line 14 to store several variables in a tuple with a particular option O, Juba but not Lee teaches parameters for options)
wherein the training for the particular one of the [parameterized] options includes storing, within the replay buffer and for a particular action from the intra-option policy for the particular one of the [parameterized] options, data… (Lee Pg. 5, Algorithm 1 instructs on line 14 to store several variables associated with a particular option O to buffer B, with other variables in the tuple being a, standing for an action, which is sampled at line 11 from intra-option policy πOo, Juba but not Lee teaches parameters for options)
data including: an initial state, the particular action, a reward… (Lee Pg. 5, Algorithm 1 instructs on line 14 to store several variables associated with a particular option O to buffer B, with other variables in the tuple being current state s, action a, and intrinsic reward ri)
and the intra-option policy for the particular one of the parameterized options is updated using the data ((Lee Pg. 5) “In the rollout phase, the HRL agent samples an option Oo using the SMDP policy µ. If the Oo was never selected before, we create the option and initialize the policy πOo and add it to a container. Next, the sample trajectories are generated by using the πOo until it terminates. After sampling one-step state transition, we compute the intrinsic reward following Definition 7. Then, the HRL agent updates the option policy and the SMDP policy using the samples stored in the buffer in the training phase”)
Konidaris further teaches:
data including:…[the particular action, a reward,] a subsequent state,… ((Konidaris Pg. 1) “A single Markov decision process is defined by a tuple <S, A, R, T, γ>, where S is a state space, A is a set of available actions, R(s, a, s’) defines the reward obtained when executing action a in state s and transitioning to state s’”, Lee also teaches a particular action and a reward)
Juba further teaches:
data including:…and one or more values associated with the one or more option parameters, ((Juba Pg. 2) “A lifted action A ∈ 𝒜 is a pair <name, params> where name is a symbol and params is a list of types, denoted name(A) and params(A), respectively, and arity(A, t) denotes the number of type-t parameters. The action model M for a set of actions A is a pair of functions preM and effM that map every action in 𝒜 to its preconditions and effects…A parameter-bound literal L for the lifted action A is a pair of the form <L, bL,A> where b is a parameter binding of L and A. preM(A) and effM(A) are sets of parameter-bound literals for A”, a set of actions corresponds to an option, a set of parameter-bound literals for action A corresponds to values for parameters for part of the option)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Lee, Konidaris, Juba, and Ng for the parent claim of claim 3, claim 2. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 6,
Lee, Konidaris, Juba, and Ng jointly teach The method of claim 1,
Lee further teaches:
wherein the mapping function is generated using planning annotated reinforcement learning (PaRL) ((Lee Pg. 3) “A PaRL task is a triple E := <M, Π, L>, where M := <S, A, P, r, s0, G, γ> is a goal-oriented MDP over RL states S, Π := <V, O, s’0, s∗> is a planning task over planning states S, and L : S → S’ is a surjective mapping from the states of the MDP S to planning task S’ satisfying s’0 = L(s0) and s∗ consistent with L(s) for all s ∈ G”, the PaRL task includes a mapping L between MDP S states and planning task S’)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Lee, Konidaris, Juba, and Ng for the parent claim of claim 6, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claim 7,
Lee, Konidaris, Juba, and Ng jointly teach The method of claim 1,
Lee further teaches:
wherein the planning domain includes a plurality of options for performing the goal ((Lee Pg. 4) “For any pair of initial state s0 ∈ S and a goal sg ∈ G in M, we can generate a sequence of options {Oo1, Oo2, ..., Ook} from a plan in Π that reaches the goal state L(sg) ∈ S’ from the initial state L(s0) ∈ S’”)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Lee, Konidaris, Juba, and Ng for the parent claim of claim 7, claim 1. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claims 8-10, 13, and 14,
Claims 8-10, 13, and 14 recite a computer hardware system for performing the function of the method of claims 1-3, 6, and 7, respectively. Specifically, claim 8 recites A computer hardware system for generating a policy for performing a goal and including a plurality of intra-option policies, comprising: a hardware processor configured to perform the following executable operations: [the method of claim 1]. Lee recites: (Lee Pg. 14) “We used two types of hardwares for evaluation. For the N-rooms and logistics domain, we used CPU only machines with 56 cores of 2 GHz CPUs and for the Montezuma’s revenge domain, we used P100 GPU for the training”.
All other limitations in claims 8-10, 13, and 14 are substantially the same as those in claims 1-3, 6, and 7, respectively, therefore the same rationale for rejection applies.
Regarding claims 15-17 and 20,
Claims 15-17 and 20 recite a computer program product comprising a computer-readable storage medium storing program code for performing the function of the method of claims 1-3 and 6, respectively. Specifically, claim 15 recites A computer program product, comprising: a computer readable storage medium having stored therein program code for generating a policy for performing a goal and including a plurality of intra-option policies, the program code, which when executed by a computer hardware system, causes the computer hardware system to perform: [the method of claim 1]. Lee recites: (Lee Pg. 14) “We implemented Algorithm 2 using python language by extending RL agents in stable-baselines 3 and using pyperplan as AI planner…For the N-rooms and logistics domain, we used CPU only machines with 56 cores of 2 GHz CPUs and for the Montezuma’s revenge domain, we used P100 GPU for the training”, python language is program code, and CPUs and GPUs inherently include computer-readable memory components.
All other limitations in claims 15-17 and 20 are substantially the same as those in claims 1-3 and 6, respectively, therefore the same rationale for rejection applies.
Claims 4, 5, 11, 12, 18, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Lee, in view of Konidaris, further in view of Juba, further in view of Ng, further in view of Horcik.
Regarding claim 4,
Lee, Konidaris, Juba, and Ng jointly teach The method of claim 1,
Lee further teaches:
wherein the MDP [distribution] defines an environment ((Lee Pg. 5) “To empirically evaluate the proposed approach, we conducted experiments on the three goal oriented MDP environments”, Konidaris but not Lee teaches a distribution for MDPs) including a plurality of MDPs meeting constraints, ((Lee Pg. 8) “We design a general method for injecting intrinsic rewards to RL agents from the abstract planning task by reformulating the underlying decomposed sub-MDPs with constraints visible to planning agents”)
Horcik teaches the following further limitation that Konidaris does not teach, and more explicitly than Lee, Juba, or Ng teaches:
and the constraints including a predicate, action, object, type, and action model ((Horcik Pg. 2) “Definition 2. A normalized PDDL task is a tuple 𝒫 = <B, T, V, P, A, ψI, ψG> where B is a non-empty set of objects, T is a type hierarchy over B…P is a set of predicate symbols…An action schema a(x-arrow) ∈ A is a tuple a = <prea(x-arrow), adda(x-arrow), dela(x-arrow)> where prea(x-arrow), adda(x-arrow), and dela(x-arrow) are sets of atoms, called preconditions, add effects, and delete effects, respectively, and x-arrow is a tuple of variables occurring in any atom in a(x-arrow). If we substitute a tuple of objects…for x, we create a ground action (or shortly just an action). The resulting ground action is denoted by a(b-arrow)”, according to Applicant’s specification at [0041] “Lifted Action Model (LAM) (also known as an action schema)”, therefore an action schema is an action model)
At the time of filing, one of ordinary skill in the art would have motivation to combine Lee, Konidaris, Juba, Ng, and Horcik by taking the method of claim 1 for training intra-option policies using mappings from MDP states, from MDPs in a distribution, and planning states, and including defining an environment with constraints with the MDP distribution, jointly taught by Lee, Konidaris, Juba, and Ng, and additionally including the constraints being specifically a predicate, a type, an action, a type, and an action model, taught by Horcik, as Horcik teaches: (Horcik Pg. 1) “Classical planning tasks are usually modeled in the standard PDDL language based on first-order logic”, with the constraints being part of the PDDL (Planning Domain Definition) language, doing so is a well-known technique for outlining a planning problem in a standardized and human-readable fashion, imparting the predictable benefit of being known and easy to work with for experts in the field, increasing the ease of performing the method. Such a combination would be obvious.
Regarding claim 5,
Lee, Konidaris, Juba, Ng, and Horcik jointly teach The method of claim 4,
Lee further teaches:
wherein the policy for performing the goal is configured to be used with the additional MDP, wherein the additional MDP meets the constraints ((Lee Pg. 8) “Our framework links the state abstraction in AI planning and temporal abstraction in RL, providing a way to decompose the MDP into sub-MDPs by specifying options initiation set and termination condition based on planning operator definitions. We design a general method for injecting intrinsic rewards to RL agents from the abstract planning task by reformulating the underlying decomposed sub-MDPs with constraints visible to planning agents. Learning only the (intra-option) policies for these sub-MDPs is shown to work well in practice on various problems, significantly improving sample efficiency”, a sub-MDP is an additional MDP)
At the time of filing, one of ordinary skill in the art would have motivation to combine the method jointly taught by Lee, Konidaris, Juba, Ng, and Horcik for the parent claim of claim 5, claim 4. No new embodiments are introduced, so the reason to combine is the same as for the parent claim.
Regarding claims 11 and 12,
Claims 11 and 12 recite a computer program product comprising a computer-readable storage medium storing instructions for performing the function of the method of claims 4 and 5, respectively. All other limitations in claims 11 and 12 are substantially the same as those in claims 4 and 5, respectively, therefore the same rationale for rejection applies.
Regarding claims 18 and 19,
Claims 18 and 19 recite a computer hardware system for performing the function of the method of claims 4 and 5, respectively. All other limitations in claims 18 and 19 are substantially the same as those in claims 4 and 5, respectively, therefore the same rationale for rejection applies.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Fiser “Lifted Fact-Alternating Mutex Groups and Pruned Grounding of Classical Planning Problems” discloses a proof of characteristics of lifted mutex groups and applies this to provide a more efficient method of translation to a grounded representation.
Sutton et al. “Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning” discloses planning with options and intra-option methods for reinforcement learning with Markov Decision Processes and Semi-Markov Decision Processes.
Xu et al. “Hierarchical Reinforcement Learning in StarCraft II with Human Expertise in Subgoals Selection” discloses a method for teaching agents to play a video game using reinforcement learning.
Rezaei-Shoshtari et al. (U.S. Patent Application Publication No. 2024/0198517) discloses a method for training a hypernetwork to provide a policy for a new task by mapping a family of parameterized MDPs to a family of near optimal policies.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VICTOR A NAULT whose telephone number is (703) 756-5745. The examiner can normally be reached M - F, 12 - 8.
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, Miranda Huang can be reached at (571) 270-7092. 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.
/V.A.N./Examiner, Art Unit 2124
/MIRANDA M HUANG/ Supervisory Patent Examiner, Art Unit 2124