Negotiation methods and systems for implementation, including the selection of neural networks.

Abstract

A system for selecting and using a neural architecture includes a non-temporary computer-readable medium configured to store instructions, and a processor connected to the non-temporary computer-readable medium. The processor is configured to execute instructions for receiving negotiation traces, including offers from previous negotiations with a target negotiating party. The processor is configured to execute instructions for training a neural architecture using only the received negotiation traces. The processor is configured to execute instructions for optimizing offers to the target negotiating party during negotiations using the trained neural architecture.

Description

Technical Field

[0001] The present disclosure relates to systems, methods, and non-transitory computer-readable media for selecting and using neural architectures.

Background Art

[0002] With the increase in commercial transactions using the Internet, autonomous negotiation is also increasing. Autonomous negotiation attempts to reach an agreement between parties by using artificial intelligence (AI) instead of human negotiators. AI is used to evaluate offers from the opponent and then determine a response to the offer. In some cases, the response is a counteroffer. In some cases, the response is acceptance of the offer. In some cases, the response is to end the negotiation.

[0003] In some approaches, the AI negotiator is fully trained with the user's preferences and priorities before the start of the negotiation. In some approaches, the AI negotiator includes some small uncertainties regarding the user's preferences and priorities at the start of the negotiation. In some approaches, the AI negotiator can submit queries to the user to reduce the uncertainty of the user's preferences and priorities. However, the query only asks how valuable a particular parameter is to the user without asking about the user's preference between options.

Summary of the Invention

[0004] According to a first exemplary aspect of the present disclosure, a system is provided for selecting and using a neural architecture, comprising a non-temporary computer-readable medium configured to store instructions, and a processor connected to the non-temporary computer-readable medium. The processor is configured to receive negotiation traces, including offers from previous negotiations with a target negotiating party; to train a neural architecture using only the received negotiation traces; and to execute instructions using the trained neural architecture for optimizing an offer to the target negotiating party during negotiations.

[0005] A second exemplary aspect of the present disclosure provides a method for selecting and using a neural architecture, which includes receiving a negotiation trace containing offers from previous negotiations with a target negotiating party, training a neural architecture using only the received negotiation trace, and using the trained neural architecture to optimize an offer to the target negotiating party during negotiations.

[0006] According to a third exemplary aspect of the present disclosure, a non-temporary computer-readable medium is provided which is configured to store instructions for selecting and using a neural architecture. The instructions are configured to cause a processor to perform an operation, which includes receiving a negotiation trace containing offers from previous negotiations with a target negotiating party, training a neural architecture using only the received negotiation trace, and optimizing an offer to the target negotiating party during negotiations using the trained neural architecture. [Brief explanation of the drawing]

[0007] The aspects of this disclosure will be best understood from the following detailed description in conjunction with the attached drawings. Note that, in accordance with standard industry practice, various features are not depicted to scale. In fact, the dimensions of various features may be increased or decreased as appropriate for clarity in the description. [Figure 1] This is a schematic diagram of negotiations in several embodiments. [Figure 2] This is a flowchart showing how to select a negotiation neural network based on several embodiments. [Figure 3] This is a schematic diagram illustrating a method for constructing a neural network according to several embodiments. [Figure 4] This is a schematic diagram illustrating a method for building and retraining a neural network according to several embodiments. [Figure 5] This is a flowchart illustrating how to select a response to an offer, according to several embodiments. [Figure 6] This is a block diagram of a system for conducting negotiations, based on several embodiments. [Modes for carrying out the invention]

[0008] The following disclosure provides many different embodiments or examples for implementing different features of the subject matter provided. To simplify this disclosure, specific examples of components, values, behaviors, materials, arrangements, etc., are described below. Naturally, these are merely examples and are not intended to be limiting. Other components, values, behaviors, materials, arrangements, etc., are possible. For example, the formation of a first feature over a second feature in the following description may include embodiments in which the first and second features are formed in direct contact, or it may include embodiments in which additional features may be formed between the first and second features so that the first and second features do not have to be in direct contact. Furthermore, this disclosure may repeat reference numbers and / or letters in various examples. This repetition is for simplification and clarity and does not in itself presuppose any relationships between the various embodiments and / or configurations discussed.

[0009] Negotiations utilizing automated systems are used to increase the speed at which solutions are reached among multiple stakeholders. Each stakeholder seeks to maximize the value or utility function of the negotiation outcome. However, each stakeholder also seeks to avoid alienating other stakeholders by avoiding offers or proposals that are completely contrary to their views. By knowing which criteria are most beneficial to other stakeholders, the negotiation system can generate the highest offers and proposals that are most likely to improve value for the user of the negotiation system, while avoiding alienating or damaging relationships with other stakeholders in the negotiation. In some cases, this type of situation is called reducing the Pareto-frontier distance.

[0010] Examining previous negotiations involving the other party, also referred to herein as a self-interested party, helps determine which criteria are evaluated by the other party. This description includes a system for generating a neural network based on past negotiations performed by the other party. In some embodiments, negotiations include offers, counter-offers, acceptances, rejections, or other negotiation procedures. In some embodiments, individual negotiation procedures are referred to as negotiation traces or transcripts. These traces can be used to generate a neural network to negotiate with the parties involved in the previous negotiations, or to adapt an existing neural network. In some embodiments, the past negotiations are within the same domain as the proposed negotiations. In some embodiments, the past negotiations are within a different domain than the proposed negotiations. As the similarity between the domains of the proposed negotiations and the past negotiations increases, the accuracy and precision of the predictions and offers generated by the neural network improve.

[0011] In some cases, a change in the business circumstances of a party can alter its priorities. In some embodiments, the system includes the ability to evaluate an ongoing negotiation to determine whether the neural network currently in use is adequately predicting offers and counter-offers from the other party. In response to a determination that the system is not adequately predicting offers and counter-offers from the other party, the system may recommend modifications to the neural network to improve its accuracy in predicting actions by the other party. In some embodiments, the system recommends a replacement neural network based on traces received during the ongoing negotiation. While the following description focuses primarily on bilateral negotiations, those skilled in the art will understand that the current application is not limited to bilateral negotiations and that other multilateral negotiations may be performed and evaluated by the system described herein.

[0012] Figure 1 is a schematic diagram of negotiation 100 using an autonomous negotiation method in several embodiments. Negotiation 100 includes a first negotiator 110, which includes a first strategy 112. Negotiation 110 includes a second negotiator 120, which includes a second strategy 122. Protocol 130 is used to determine the style of negotiation between the first negotiator 110 and the second negotiator 120. Domain 140 includes the issues to be resolved in negotiation 100. Negotiation 100 is a bilateral negotiation. However, this disclosure is not limited to bilateral scenarios.

[0013] The first negotiator 110 includes an AI algorithm for determining the priorities of the issues to be resolved. These priorities are incorporated into utility values ​​associated with each of the issues to be resolved. The first negotiator 110 uses these utility values ​​to prioritize the possible outcomes from negotiation 100. The first negotiator 110 can communicate with the second negotiator 120 using protocol 130. In some embodiments, the first negotiator 110 can communicate wirelessly. In some embodiments, the first negotiator 110 can communicate using a wired connection. In some embodiments, the first negotiator 110 can communicate over the Internet. The first negotiator 110 can send offers, counteroffers, acceptances, or rejections to the second negotiator 120. The first negotiator 110 is autonomous, i.e., operates without user input or control.

[0014] The first strategy 112 is an action determined by the first negotiator 110 based on the utility values ​​and algorithm used by the first negotiator 110. In some embodiments, the first strategy 112 includes an offer, a counter-offer, an acceptance, or a rejection. In some embodiments, the first strategy 112 changes during negotiation 100 based on new information received from the second negotiator 120.

[0015] The second negotiator 120 includes an AI algorithm for determining priorities acquired using utility values ​​for the issues to be resolved. The second negotiator 120 can communicate with the first negotiator 110 using protocol 130. In some embodiments, the second negotiator 120 can communicate wirelessly. In some embodiments, the second negotiator 120 can communicate using a wired connection. In some embodiments, the second negotiator 120 can communicate over the Internet. The second negotiator 120 can send offers, counteroffers, acceptances, or rejections to the first negotiator 110. In some embodiments, the hardware of the first negotiator 110 is the same as the hardware of the second negotiator 120. In some embodiments, the hardware of the first negotiator 110 is different from the hardware of the second negotiator 120. In some embodiments, the algorithm implemented in the first negotiator 110 is the same as the algorithm implemented in the second negotiator 120. In some embodiments, the algorithm implemented in the first negotiator 110 is different from the algorithm implemented in the second negotiator 120. In some embodiments, the second negotiator 120 is autonomous. In some embodiments, the second negotiator 120 is controlled based on user input.

[0016] The second strategy 122 is an action determined by the second negotiator 120 based on the utility values ​​and algorithms used by the second negotiator 120. In some embodiments, the second strategy 122 includes an offer, a counter-offer, an acceptance, or a rejection. In some embodiments, the second strategy 122 changes during negotiation 100 based on new information received from the first negotiator 110.

[0017] Protocol 130 is a set of rules for negotiation 100. For example, in an alternating offer protocol, negotiators, e.g., a first negotiator 110 and a second negotiator 120, exchange offers alternately. In some embodiments, protocol 130 is an alternating offer protocol. In some embodiments, protocol 130 is a different protocol, such as a time-based offer protocol, or another suitable protocol. Protocol 130 is established before negotiation 100 begins, so that the first negotiator 110 and the second negotiator 120 know when and whether an offer or response should be sent. In some embodiments, protocol 130 includes a maximum number of steps, e.g., an offer. In some embodiments, protocol 130 includes a time limit for negotiation 100.

[0018] Domain 140 includes a problem to be solved. For example, in some embodiments, the problem to be solved includes price, quantity, brand name, etc. In some embodiments, Domain 140 includes a single problem. In some embodiments, Domain 140 includes multiple problems. Each of the first negotiator 110 and the second negotiator 120 is aware of Domain 140 and assigns utility values ​​to combinations of problems to be solved. For example, if Domain 140 includes price and brand name, and there are two choices for each of price and brand name, the first negotiator 110 has four different utility values, one for each possible combination. In some embodiments, the first negotiator 110 and the second negotiator 120 have utility values ​​for fewer combinations than all possible combinations. In some embodiments, the first and second negotiators have different priorities, so the utility value assigned by the first negotiator 110 is different from the utility value assigned by the second negotiator 120.

[0019] Figure 2 is a flowchart of Method 200 for selecting a negotiation neural network according to several embodiments. In some embodiments, Method 200 can be used to create a new neural network based on target negotiating parties. In some embodiments, Method 200 can be used to select a neural network from existing neural networks based on target negotiating parties. In some embodiments, Method 200 can be used to update a neural network during negotiation. In some embodiments, Method 200 can be used to switch between neural networks during negotiation. Those skilled in the art will recognize that Method 200 can be used to perform any of the above uses individually or in combination. The target negotiating parties are the parties with whom the user of the neural network intends to negotiate. In some embodiments, Method 200 can be used to select or modify a neural network to perform negotiation 100 (Figure 1). In some embodiments, Method 200 can be used to select or modify a neural network to perform a different negotiation than negotiation 100 (Figure 1).

[0020] In operation 205, a trace is collected. The trace is the outcome of previous negotiations with the target negotiating party. In some embodiments, the trace includes offers, counter-offers, acceptances, rejections, or other appropriate responses during negotiations. The trace includes the terms that were ultimately accepted, but in some embodiments, the trace also includes intermediate exchanges during negotiations. In some embodiments, the trace is collected from stored data in a database based on previous automated negotiations with the target negotiating party. In some embodiments, the trace is collected from other sources, such as emails, transcripts, scanned documents, or other sources of appropriate exchanges, during previous negotiations with the target negotiating party. By limiting the trace to the target negotiating party, method 200 can arrive at a neural network that focuses on the priority of the target negotiating party instead of a more general neural network that could be used in negotiations with other parties. As a result, method 200 can produce higher value negotiation outcomes when negotiating with the target negotiating party compared to other general approaches.

[0021] In operation 210, an initial architecture is specified. The initial architecture can be used to determine the structure of the neural network and to train the neural network to negotiate with the target negotiating party. In some embodiments, the neural network can be used to predict the opponent's second strategy 122 (Figure 1) in negotiations. In some embodiments, the details of operation 210 are described with respect to method 300 (Figure 3), which will be discussed below. In some embodiments, a single initial architecture is defined. In some embodiments, multiple initial architectures are defined to provide various options during future negotiations. In some embodiments, the initial architecture defines parameters of the neural network, such as depth, width, specifiers, or other appropriate parameters. The depth of the neural network specifies the number of layers in the neural network. As the number of layers increases, the accuracy of the neural network increases. However, the processing power used to implement the neural network also increases. The width of the neural network determines how many nodes exist within each layer of the neural network. As the number of nodes per layer increases, the accuracy of the neural network increases. However, the processing power used to implement the neural network also increases. The specifier determines the type of architecture used to form the neural network. In some embodiments, the specifier is selected from long short-term memory (LSTM), gated recurrent units (GRU), transformer, convoluted neural network (CNN), fully connected layers (FCL), or other suitable types of architecture. In some embodiments, each of the neural networks generated in operation 210 contains the same initial architecture but has different node weights after training.In some embodiments, at least one neural network generated in operation 210 has an architecture different from at least one other neural network. Utilizing different architectures helps provide different options for utilizing neural networks during future negotiations with target negotiation parties.

[0022] In some embodiments, input parameters are utilized to help train one or more initial architectures. In some embodiments, input parameters include channels, types, cardinalities, or other suitable input parameters. The input parameters are fed into the initial architecture, and the output of the initial architecture is compared with the traces collected in operation 205 to help train the initial architecture using the collected traces.

[0023] The channel contains the value of each offer or counter-offer within the negotiation. In some embodiments, the channel contains the outcome of the offer or counter-offer, a self-utility function, a predicted counterparty utility function, context, time, round, or other appropriate parameters. Those skilled in the art will understand that the above parameters of the channel can be used individually or in any combination. The outcome shows the counter-offer's response to the offer. The self-utility function determines the value of the offer or counter-offer to the user of the neural network. The predicted counterparty utility function is the predicted value of the offer or counter-offer to the target negotiating party. The context includes additional available information about the target negotiating party, such as stock prices, economic performance, or other available information. The time indicates the time within the negotiation when the offer or counter-offer was made. Some negotiations are set to terminate after a certain period if acceptance is not reached before the end of the period. The time parameter indicates at what point in the period the offer or counter-offer was made. The round indicates the round in which the offer or counter-offer was made. Some negotiations are set to terminate after a certain number of rounds of offers and counter-offers if acceptance is not reached. The round parameter indicates at what stage of negotiations an offer or counter-offer was made.

[0024] The type parameter includes information on what information is being utilized in the channel. In some embodiments, the type parameter includes the overall history, the self-policy, the opponent policy, or other suitable types of information. The overall history includes the results of both the user of the neural network and the target negotiation party. The self-policy includes information on how the user of the neural network responded during the trace. The opponent policy includes information on how the target negotiation party responded during the trace. Because the responses of both parties are considered, the overall history provides a more accurate and precise trained neural network. However, the increased amount of data provided to train the initial architecture increases the processing time for training the neural network.

[0025] The cardinality parameter includes how many offers or counter-offers are input into the initial architecture for training the neural architecture. Cardinality can range from the portion of the available data selected by the user to all of the available data. As the number of offer and counter-offer inputs for training the initial architecture increases, the accuracy and precision of the neural network increases. However, the increased amount of data provided to train the initial architecture also increases the processing time for training the neural network.

[0026] In operation 215, an algorithm is applied to select which of the trained architectures from operation 210 to utilize in negotiations with the target negotiation party. In some embodiments, the algorithm includes negotiating one or more trained architectures with each other. In some embodiments, the algorithm includes causing the trained architecture to attempt to replicate previous negotiations. In some embodiments, if a single trained architecture is generated in operation 210, operation 215 is omitted.

[0027] In some embodiments, the selection of a trained architecture is based on an evaluation of the prediction of the target negotiating party's actions, as well as the prediction of the neural network user's response to the target negotiating party's actions. In some embodiments, the selection is based on which trained architecture has the best accuracy in predicting the target negotiating party's actions and / or the neural network user's response. In some embodiments, the selection is based on which trained architecture produces the best value outcome for the neural network user.

[0028] In operation 220, the selected trained architecture is used in negotiations with the target negotiating party. In some embodiments, operation 220 includes negotiations similar to negotiation 100 (Figure 1). During each exchange of offers or counteroffers in operation 220, traces are provided back to operation 205 for the generation of a new architecture or the retraining of an existing architecture.

[0029] In some embodiments, operation 225 is performed following operation 220. In some embodiments, operation 225 is performed after all offers or counteroffers received from the target negotiating party. In some embodiments, method 200 proceeds to operation 225 following a predetermined progression by negotiation. In some embodiments, the predetermined progression is determined based on user input to the neural network. In some embodiments, the predetermined progression includes a ratio of the elapsed negotiation time to the maximum time allowed for negotiation. In some embodiments, the predetermined progression includes a ratio of the elapsed rounds of negotiation to the maximum number of rounds allowed for negotiation.

[0030] In operation 225, the performance of the trained architecture used in operation 220 is evaluated to determine whether to continue using the same trained architecture as the negotiation progresses, or to switch to a different trained architecture for subsequent steps in the negotiation. In some embodiments, operation 225 is performed only once during the negotiation. In some embodiments, operation 225 is performed multiple times during the negotiation. In some embodiments, the decision on whether the trained architecture accurately predicts the steps of the negotiation is based on the difference between the offer or counter-offer received from the target negotiating party and the offer or counter-offer predicted by the trained architecture. In some embodiments, the difference is based on the neural network's value to the user for the offer or counter-offer. In some embodiments, the difference is based on the predicted values ​​for the target negotiating party for the offer and counter-offer. In some embodiments, accuracy is determined based on a threshold to account for small variability in the negotiation with the target negotiating party. In some embodiments, the threshold is adjusted as the negotiation progresses. For example, in some embodiments, the threshold decreases as the negotiation progresses, meaning that inaccuracies by the trained architecture used in operation 220 are acceptable in the subsequent negotiation.

[0031] Utilizing action 225 helps improve the value of the neural network to the user in situations where method 200 adjusts the negotiation strategy compared to past negotiations. For example, in some cases, as the size of the target negotiation parties increases, the speed of delivery of the goods becomes more important than the price of the goods. As a result, the traces used to train the architecture in action 210 are not very accurate in predicting the actions of the target negotiation parties during negotiation in action 220. Therefore, adjusting the user's negotiation strategy helps improve the value of the negotiation to the user.

[0032] In response to a decision that the accuracy of the trained architecture used in operation 220 is satisfactory, method 200 returns to operation 220, and the negotiation continues using the previously selected trained architecture. In response to a decision that the accuracy of the trained architecture used in operation 220 is insufficient, method 200 returns to operation 215, and a different trained architecture from the one previously used in operation 220 is selected for the current negotiation. In some embodiments, method 200 prevents operation 215 from selecting any trained architecture previously used in the ongoing negotiation in subsequent iterations of operation 215. By preventing the selection of a previously used trained architecture that has already proven insufficient for the current negotiation, method 200 improves the likelihood of a higher-value outcome for the user of the neural network.

[0033] Those skilled in the art will recognize that modifications of Method 200 are within the scope of this description. In some embodiments, the order of operations of Method 200 is changed. For example, in some embodiments, the structure of the neural network architecture is defined before collecting traces for use in training the architecture. In some embodiments, additional operations are included in Method 200. For example, in some embodiments, the user is prompted to input criteria for determining the satisfactory accuracy of a selected trained architecture. In some embodiments, at least one operation of Method 200 is omitted. For example, in some embodiments that include only a single trained architecture, operation 225 is omitted.

[0034] Method 200 utilizes only traces collected to train a neural network architecture. As a result, the trained architecture aims to improve value for the user when negotiating with a target negotiating party. This approach is particularly advantageous for the user in situations where the user frequently negotiates with a target negotiating party. Furthermore, Method 200 provides feedback during future or ongoing negotiations for the development of new architectures or the retraining of existing architectures. As a result, value for the user is more likely to improve during future negotiations with the target negotiating party. In addition, Method 200 provides automatic switching between trained architectures to account for potential adjustments to the target negotiating party's negotiation strategy compared to previous negotiations. This enables Method 200 to help maintain or improve the value of negotiations as the needs and desires of the target negotiating party evolve over time.

[0035] Figure 3 is a schematic diagram of method 300 for constructing a neural network according to several embodiments. In some embodiments, method 300 can be used to perform operation 210 of method 200 (Figure 2). In some embodiments, method 300 can be used to develop a negotiator, e.g., negotiator 110, in negotiation 100 (Figure 1). In some embodiments, method 300 can be used to develop and train a neural network usable in methods other than method 200 (Figure 2) or in negotiation 100 (Figure 1).

[0036] Method 300 includes inputting input parameters into a neural architecture using an input adapter in operation 310. The input parameters are used to help train one or more initial architectures. In some embodiments, the input parameters include channels, types, cardinality, or other appropriate input parameters. Examples of channel, type, and cardinality parameters have been discussed above and will not be repeated in detail here for brevity. In some embodiments, the values ​​of the input parameters are based on available traces of the target negotiating party. In some embodiments, the user can select from available traces to determine the input parameters. In some embodiments, the user selects several input parameters to use to train the neural architecture, and a system such as system 600 (Figure 6) is used to select traces to use as input parameters based on the input received from the user. In some embodiments, the most recent trace that satisfies the user input is selected as an input parameter. In some embodiments, traces that satisfy the user input are randomly selected as input parameters.

[0037] Method 300 further includes generating a neural architecture in operation 320. In some embodiments, a single initial architecture is defined. In some embodiments, multiple initial architectures are defined to provide various options during future negotiations. In some embodiments, the initial architecture defines parameters of the neural network, such as depth, width, specifiers, or other appropriate parameters. The depth of the neural network specifies the number of layers in the neural network. As the number of layers increases, the accuracy of the neural network increases. However, the processing power used to implement the neural network also increases. The width of the neural network determines how many nodes exist within each layer of the neural network. As the number of nodes per layer increases, the accuracy of the neural network increases. However, the processing power used to implement the neural network also increases. The specifiers determine the type of architecture used to form the neural network. In some embodiments, the specifiers are selected from LSTM, GRU, transformer, CNN, FCL, or other appropriate types of architecture. In some embodiments, each of the neural networks generated in operation 320 contains the same initial architecture but has different node weights after training. In some embodiments, at least one neural network generated in operation 320 has a different architecture from at least one other neural network. Utilizing different architectures helps provide different options for utilizing the neural network during future negotiations with target negotiating parties.

[0038] Method 300 further includes receiving output parameters using an output adapter 330. The output parameters are output from the neural architecture. By comparing the output parameters with actual data from the trace, the accuracy of the neural architecture can be determined, and the weights within the nodes of the neural architecture can be adjusted as part of the training process of the neural architecture.

[0039] Those skilled in the art will recognize that Method 300 can train multiple different neural architectures. Training more neural architectures provides additional options for use during negotiation, such as being able to switch between trained architectures during negotiation. However, training more neural architectures increases the processing load on the system used to train the neural architectures.

[0040] Method 300 utilizes only traces from previous negotiations with the target negotiating party to train the neural network architecture. As a result, the trained architecture aims to improve the value for the user when negotiating with the target negotiating party. This approach is particularly advantageous for the user in situations where the user frequently negotiates with the target negotiating party.

[0041] Figure 4 is a schematic diagram of Method 400 for constructing and retraining a neural network, according to several embodiments. In some embodiments, Method 400 is used in combination with Method 200 (Figure 2) or Method 300 (Figure 3) to generate one or more trained neural architectures for performing negotiations such as Negotiation 100 (Figure 1). In some embodiments, Method 400 can be used in ways other than Method 200 (Figure 2) or Method 300 (Figure 3) to generate and select neural architectures for negotiations other than Negotiation 100 (Figure 1). Method 400 includes receiving an initial set of data and repeatedly retraining and updating the neural architecture to help improve its value in future negotiations with target negotiating parties.

[0042] In operation 405, an initial population of trained neural architectures is received. In some embodiments, the initial population includes a single trained neural architecture. In some embodiments, the initial population includes multiple trained neural architectures. In some embodiments, the initial population is generated using method 300 (Figure 3). In some embodiments, the initial population is generated using a method other than method 300 (Figure 3). In some embodiments, the initial population is stored in a database for selection and retraining during subsequent iterations of method 400.

[0043] In operation 410, the next generation of trained neural architectures is stored in a database for future negotiated selections. In the first iteration of method 400, the next generation is equal to the initial population. In some iterations, at least one trained neural architecture from a previous iteration is not updated by method 400.

[0044] In operation 415, a decision is made regarding whether to mutate one or more of the trained neural architectures. Mutation allows the architecture to adjust the weights of one or more nodes while simultaneously adjusting the structure of the architecture. In some embodiments, a decision is made to adjust the width or depth of the neural architecture in response to the inability to converge the neural architecture with sufficient accuracy.

[0045] In operation 420, an architecture crossover is performed. An architecture crossover occurs when one or more nodes from a first architecture are imported into a second architecture. In some cases, the crossover helps to avoid premature convergence in situations where the architectures are not sufficiently accurate. In other cases, the crossover helps to improve the accuracy of the second architecture in situations where one or more nodes in the second architecture are identified as a potential source of inaccuracy.

[0046] In operation 425, the mutated and / or crossover architecture is trained. Architecture training is performed using traces from the target negotiating party to help improve the value of future negotiation outcomes with the target negotiating party. In some embodiments, training is performed using a weighted loss function or another suitable loss function.

[0047] In operation 430, the selection of a trained architecture is performed to determine which trained architecture will be used at the start of the next negotiation with the target negotiating party. In some embodiments, the trained architectures are switched during the negotiation, as discussed above with respect to method 200 (Figure 2). In some embodiments, the selection is performed by making the trained architectures compete against each other, either based on a trace or based on a simulated negotiation with the target negotiating party.

[0048] Method 400 proceeds from operation 430 to operation 410. In some embodiments, the trained architecture selected from operation 430 is identified in the next generation to assist in mutation, crossover, and selection during subsequent iterations of Method 400.

[0049] Method 400 utilizes only traces from previous negotiations with the target negotiating party to train the neural network architecture. As a result, the trained architecture aims to improve the value for the user when negotiating with the target negotiating party. This approach is particularly advantageous for the user in situations where the user frequently negotiates with the target negotiating party.

[0050] Figure 5 is a flowchart of Method 500 for selecting a response to an offer, according to several embodiments. Method 500 includes multiple trained architectures or models. Those skilled in the art will understand that in some embodiments, Method 500 relies on a single trained architecture or model. In some embodiments, Method 500 is used in combination with Method 200 (Figure 2), Method 300 (Figure 3), or Method 400 (Figure 4) to determine an offer as part of a negotiation, such as Negotiation 100 (Figure 1). In some embodiments, Method 500 can be used in ways other than Method 200 (Figure 2), Method 300 (Figure 3), or Method 400 (Figure 4) to determine an offer for a negotiation other than Negotiation 100 (Figure 1).

[0051] In operation 505, a negotiation trace is received. In some embodiments, the trace includes offers, counteroffers, acceptances, rejections, or other appropriate responses during negotiations. The trace includes the terms that were ultimately accepted, but in some embodiments, the trace also includes intermediate exchanges during negotiations. In some embodiments, the trace is collected from stored data in a database based on previous automated negotiations with the target negotiating party. In some embodiments, the trace is collected from other sources, such as emails, transcripts, scanned documents, or other sources of appropriate exchanges, during previous negotiations with the target negotiating party. By limiting the trace to the target negotiating party, Method 500 can reach, select, and utilize a neural network focused on the priority of the target negotiating party, instead of a more general neural network available for negotiations with other parties. As a result, Method 500 can produce higher value negotiation outcomes when negotiating with the target negotiating party compared to other general approaches.

[0052] In operation 510, the available trained models are evaluated to determine which trained model to use in negotiation. In some embodiments, which include a single trained model, operation 510 is skipped. In some embodiments, multiple models are selected, and the selected multiple models “vote” on the offer in operation 515, which will be discussed later. An algorithm is used to evaluate the models. In some embodiments, the algorithm includes having one or more trained architectures negotiate with each other. In some embodiments, the algorithm includes having the trained architectures attempt to replicate previous negotiations.

[0053] In some embodiments, the selection of a trained architecture is based on an evaluation of the prediction of the target negotiating party's actions, as well as the prediction of the neural network user's response to the target negotiating party's actions. In some embodiments, the selection is based on which trained architecture has the best accuracy in predicting the target negotiating party's actions and / or the neural network user's response. In some embodiments, the selection is based on which trained architecture produces the best value outcome for the neural network user.

[0054] In operation 515, the selected model or multiple models vote for the next offer to be sent to the target negotiating party. In embodiments including a single trained model, the single trained model selects the next offer to be sent to the target negotiating party. In some embodiments including multiple selected trained models, each of the selected trained models proposes an offer that includes one or more parameters. A sum is determined for each parameter across the models. The offer with the parameter that has the highest sum is determined to be the winner of the vote. The output of the vote is used in operation 535, which will be discussed later.

[0055] In operation 520, a decision is made as to whether a predicted counterpart model is available. The predicted counterpart model is a prediction of a neural network to be used by the target negotiating party. In some embodiments, the predicted counterpart model is generated using, for example, method 300 (Figure 3) based on the negotiation trace received in operation 505. In response to the decision that a predicted counterpart model is available, method 500 proceeds to operation 525. In response to the decision that no predicted counterpart model is available, method 500 proceeds to operation 535.

[0056] In operation 525, the predicted peer model is evaluated. Operation 525 is similar to operation 510, except that one or more models being evaluated are the predicted peer model or more models. In some embodiments where a single predicted peer model is available, operation 525 is skipped.

[0057] In action 530, an offer that maximizes the counterparty's utility is determined using one or more predicted counterparty models. Maximizing the counterparty's utility means maximizing the predicted value for the counterparty based on the predicted counterparty models. Action 530 is the same as action 515, except that the offer selection is based on the predicted counterparty models.

[0058] In action 535, the best offer is selected. In embodiments where a predicted counterparty model is unavailable, action 535 selects an offer from action 515. In some embodiments where a predicted counterparty model is available, action 535 compares the offer from action 515 with the offer from action 530. In some embodiments, action 535 selects the offer received from action 515 in response that the negotiation is below a threshold period during negotiation. In some embodiments, action 535 selects the offer received from action 530 in response that the negotiation has entered a state beyond a threshold period during negotiation. That is, action 535 selects the offer that is most beneficial to the user during the initial part of the negotiation. Action 535 selects the offer that is predicted to be most valuable to the counterparty near the end of the negotiation. This approach allows method 500 to attempt to obtain the best value for the user while minimizing the risk of failing to reach an agreement during negotiation. In some embodiments, the threshold is about 95% of the maximum period during negotiation, e.g., round or time. In some embodiments, an offer received from operation 530 is rejected in response to a decision, based on input received from the user, that one or more of the offer's parameters are not accepted by the user.

[0059] In operation 540, the offer is sent to the other party and the trace is updated. The updated trace is fed back to operation 505 to determine the next offer to the other party, to generate additional architectures, or to retrain one or more existing architectures. Method 500 is repeated until an agreement is reached in the negotiation or until the negotiation period, such as a round or time, expires.

[0060] Those skilled in the art will recognize that modifications of Method 500 are within the scope of this description. In some embodiments, the order of operations of Method 500 is adjusted. For example, in some embodiments, the evaluation of the neural model is performed before negotiation begins. In some embodiments, additional operations are included in Method 500. For example, in some embodiments, a decision is made during negotiation regarding whether to switch to a different trained model for future offer decisions. In some embodiments, at least one operation of Method 500 is omitted. For example, in some embodiments that include only a single trained architecture, operation 510 is omitted.

[0061] Method 500 utilizes only the traces collected to evaluate the neural network model. As a result, the selected model aims to improve value for the user when negotiating with the target negotiating party. This approach is particularly advantageous for the user in situations where the user frequently negotiates with the target negotiating party. Furthermore, Method 500 provides feedback during future or ongoing offer decisions to develop and re-evaluate the model. As a result, value for the user is more likely to improve during negotiations with the target negotiating party.

[0062] Figure 6 is a block diagram of a system 600 for conducting negotiations in several embodiments. The system 600 includes a hardware processor 602 and a non-temporary computer-readable storage medium 604 that stores computer program code 606, i.e., a set of executable instructions. The computer-readable storage medium 604 is also encoded with instructions 607 for interfacing with external devices such as mobile terminals. The processor 602 is electrically coupled to the computer-readable storage medium 604 via a bus 608. The processor 602 is also electrically coupled to an I / O interface 610 via the bus 608. A network interface 612 is also electrically connected to the processor 602 via the bus 608. The network interface 612 is connected to a network 614 so that the processor 602 and the computer-readable storage medium 604 can connect to external elements via the network 614. The processor 602 is configured to execute computer program code 606 encoded in a computer-readable storage medium 604 in order to make the system 600 available to perform some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5).

[0063] In some embodiments, the processor 602 is a central processing unit (CPU), a multiprocessor, a distributed processing system, an application-specific integrated circuit (ASIC), and / or a suitable processing unit.

[0064] In some embodiments, the computer-readable storage medium 604 is an electronic, magnetic, optical, electromagnetic, infrared, and / or semiconductor system (or apparatus or device). For example, the computer-readable storage medium 604 includes semiconductor or solid-state memory, magnetic tape, removable computer diskette, random access memory (RAM), read-only memory (ROM), rigid magnetic disk, and / or optical disk. In some embodiments using optical disks, the computer-readable storage medium 604 includes compact disc read-only memory (CD-ROM), compact disc read / write (CD-R / W), and / or digital video disc (DVD).

[0065] In some embodiments, the storage medium 604 stores computer program code 606 configured to cause the system 600 to perform some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5). In some embodiments, the storage medium 604 also stores information necessary to perform some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5), as well as information generated while performing some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5), such as trace parameters 616, architecture parameters 618, channel parameters 620, cardinality parameters 622, type parameters 624, and / or a set of executable instructions for performing some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5).

[0066] In some embodiments, the storage medium 604 stores instructions 607 for interfacing with external components. The instructions 607 enable the processor 602 to generate instructions readable by the external components in order to effectively perform some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5).

[0067] The system 600 includes an I / O interface 610, which is coupled to external circuitry. In some embodiments, the I / O interface 610 includes a keyboard, keypad, mouse, trackball, trackpad, and / or cursor directional keys for communicating information and commands to the processor 602.

[0068] System 600 also includes a network interface 612 coupled to processor 602. The network interface 612 enables system 600 to communicate with a network 614 to which one or more other computer systems are connected. The network interface 612 includes wireless network interfaces such as BLUETOOTH®, WIFI, WiMAX, GPRS, or WCDMA®, or wired network interfaces such as ETHERNET, USB, or IEEE-1394. In some embodiments, some or all of the operations described in Method 200 (Figure 2), Method 300 (Figure 3), Method 400 (Figure 4), or Method 500 (Figure 5) are performed in two or more systems 600, and information is exchanged between different systems 600 via the network 614.

[0069] The following non-limiting examples are provided to help understand the implications of the current description in implementation. Those skilled in the art will understand that this application is not limited to these examples and that additional forms of implementation are possible using the current description. In the automated negotiation example, the hospital and the patient determine the date and time of an appointment in which the patient will visit the hospital in person or in effect. In such a situation, the negotiation trace example includes historical data of scheduled appointments and previous appointment requests. Scheduled appointments are available from patient records or other suitable sources. Previous appointment requests are available from voicemail, email, other electronic requests, notes in patient files, or other suitable sources. The negotiation trace also includes, in some embodiments, the availability of one or more physicians in the hospital or physician specialty. Availability is available based on previous appointments, set work schedules, vacation times, and other suitable sources. Physician specialty is available from career information or other suitable sources for each physician. In some embodiments, the negotiation trace further includes requests from the patient to a specific physician from one or more physicians.

[0070] One or more neural networks are designed based on network parameters determined, for example, by the hospital. One or more neural networks are trained using negotiation traces that show previous interactions between the patient and the hospital. Once one or more neural networks are trained, the system selects a neural network to be used in automated negotiations to set the patient's appointment time based on a request from the patient. In some embodiments, the request includes information such as the reason for the requested appointment, the urgency of the appointment, a request for a specific doctor, or other relevant information.

[0071] Next, the selected neural network performs automated negotiation to optimize patient appointments to meet the patient's requests while adhering to physician availability at the hospital. During automated negotiation, the selected neural network determines which appointment times to offer the patient based on the negotiation trace and the requests submitted by the patient. In some embodiments, during negotiation, the selected neural network is replaced by another neural network in response to the selected neural network exhibiting insufficient performance. At the end of automated negotiation, the neural network either offers the patient an acceptable appointment time, or the patient may reject all offered appointments and request an appointment at another hospital. In some embodiments, accepted appointment times are automatically recorded in the physician availability corresponding to that appointment to improve the accuracy of future patient appointment scheduling. In some embodiments, the system provides recommendations for alternative hospitals in response to all offered appointments being rejected.

[0072] Note 1

[0073] A system for selecting and using a neural architecture includes a non-temporary computer-readable medium configured to store instructions, and a processor connected to the non-temporary computer-readable medium. The processor is configured to execute instructions for receiving negotiation traces, which include offers from previous negotiations with the target negotiating party. The processor is configured to execute instructions for training the neural architecture using only the received negotiation traces. The processor is configured to execute instructions for optimizing the offer to the target negotiating party during negotiations using the trained neural architecture.

[0074] Note 2

[0075] The system of Appendix 1, wherein the processor is further configured to execute instructions for generating a neural architecture based on received architectural parameters, the received architectural parameters include an architecture width and an architecture depth, and training the neural architecture includes training the neural architecture generated based on the received architectural parameters.

[0076] Note 3

[0077] A system of Note 1 or 2 in which a processor is configured to execute instructions for training multiple neural architectures, and the multiple neural architectures include the neural architectures.

[0078] Note 4

[0079] A system according to any of the appendices 1 to 3, wherein the processor is configured to execute instructions for selecting the optimal neural network from among multiple neural architectures based on simulated negotiations among multiple neural architectures.

[0080] Note 5

[0081] A system of any of the above, described in Appendix 1 to 4, in which the processor is configured to execute instructions during negotiation to determine whether the accuracy of the trained neural architecture is satisfactory.

[0082] Note 6

[0083] A system according to any of Appendix 1 to 5, wherein the processor is configured to execute instructions to optimize the next offer after the previous offer, using a different trained neural architecture, in response to a decision that the accuracy of the trained neural architecture is insufficient.

[0084] Appendix 7

[0085] A system according to any of the appendices 1 to 6, wherein the processor is configured to use the trained neural architecture to execute instructions to optimize the next offer after the previous offer, in response to a decision that the accuracy of the trained neural architecture is satisfactory.

[0086] Note 8

[0087] The method for selecting and using a neural architecture includes receiving negotiation traces, including offers from previous negotiations with a target negotiating party. The method further includes training the neural architecture using only the received negotiation traces. The method further includes optimizing the offer to the target negotiating party during negotiations using the trained neural architecture.

[0088] Note 9

[0089] The method of Appendix 8, further comprising generating a neural architecture based on received architectural parameters, wherein the received architectural parameters include architecture width and architecture depth, and training the neural architecture comprises training the neural architecture generated based on the received architectural parameters.

[0090] Note 10

[0091] The method described in Appendix 8 or 9 further includes training multiple neural architectures, wherein the multiple neural architectures include the neural architecture.

[0092] Note 11

[0093] Any of the methods described in Appendix 8 to 10, further including selecting the optimal neural network from among multiple neural architectures based on simulated negotiations among multiple neural architectures.

[0094] Note 12

[0095] Any of the methods described in Appendix 8 to 11, further including determining during negotiations whether the accuracy of the trained neural architecture is satisfactory.

[0096] Note 13

[0097] In response to a decision that the accuracy of a pre-trained neural architecture is insufficient, any of the methods described in Appendix 8 to 12 further include optimizing the next offer after the previous offer using a different pre-trained neural architecture.

[0098] Note 14

[0099] In response to the determination that the accuracy of the trained neural architecture is satisfactory, any of the methods described in Appendix 8 to 13 further include using the trained neural architecture to optimize the next offer after the previous offer.

[0100] Note 15

[0101] A non-temporary, computer-readable medium configured to store instructions for selecting and using a neural architecture. The instructions are configured to cause the processor to perform actions including receiving negotiation traces containing offers from previous negotiations with a target negotiating party. The instructions are configured to cause the processor to perform actions including training a neural architecture using only the received negotiation traces. The instructions are configured to cause the processor to perform actions including optimizing an offer to a target negotiating party during negotiations using the trained neural architecture.

[0102] Note 16

[0103] The operation further includes generating a neural architecture based on received architectural parameters, the received architectural parameters including architecture width and architecture depth, and training the neural architecture includes training the neural architecture generated based on the received architectural parameters, in the non-temporary computer-readable medium of Appendix 15.

[0104] Note 17

[0105] A non-temporary computer-readable medium as specified in Appendix 15 or 16, wherein the operation involves training multiple neural architectures, the multiple neural architectures comprising training a neural architecture and selecting a neural architecture from among the multiple neural architectures as the optimal neural network based on simulated negotiations among the multiple neural architectures.

[0106] Note 18

[0107] The operation further includes determining whether the accuracy of the trained neural architecture is satisfactory during negotiation, in any non-temporary computer-readable medium as specified in Appendix 15 to 17.

[0108] Note 19

[0109] The operation further includes optimizing the next offer after the previous offer using a different trained neural architecture in response to a decision that the accuracy of the trained neural architecture is insufficient, in any non-temporary computer-readable medium as described in Appendix 15 to 18.

[0110] Note 20

[0111] The operation further includes using the trained neural architecture to optimize the next offer after the previous offer in response to a determination that the accuracy of the trained neural architecture is satisfactory, in any non-temporary computer-readable medium as described in Appendix 15 to 19.

[0112] The foregoing outlines some features of embodiments so that those skilled in the art may better understand aspects of the present disclosure. Those skilled in the art will understand that the present disclosure can be readily used as a basis for designing or modifying other processes and structures to perform the same purposes and / or achieve the same advantages as the embodiments introduced herein. Those skilled in the art should also understand that such equivalent configurations do not deviate from the spirit and scope of the present disclosure, and that various changes, substitutions, and modifications may be made herein without departing from the spirit and scope of the present disclosure.

[0113] This application claims priority based on U.S. Patent Application No. 18 / 317,112, filed on 15 May 2023, and its disclosure is incorporated herein by reference in its entirety.

Claims

1. A system for selecting and using neural architectures, A non-temporary computer-readable medium configured to store instructions, A processor connected to the aforementioned non-temporary computer-readable medium and The processor is equipped with, Receiving negotiation traces, including offers from previous negotiations with the target negotiating party, Training the neural architecture using only the received negotiation traces, and Optimizing the offer to the target negotiating party during negotiations using the aforementioned trained neural architecture. Configured to execute instructions for, system.

2. The processor generates the neural architecture based on the received architecture parameters. Further configured to execute the instructions for which the received architecture parameters include architecture width and architecture depth, and training the neural architecture includes training the neural architecture generated based on the received architecture parameters, The system according to claim 1.

3. The system according to claim 1 or 2, wherein the processor is configured to execute instructions for training a plurality of neural architectures, and the plurality of neural architectures include the neural architectures.

4. The system according to claim 3, wherein the processor is configured to execute instructions for selecting the neural architecture as the optimal neural network from among the plurality of neural architectures based on a simulated negotiation among the plurality of neural architectures.

5. The system according to any one of claims 1 to 4, wherein the processor is configured to execute instructions for determining whether the accuracy of the trained neural architecture is satisfactory during negotiation.

6. In response to a determination that the accuracy of the trained neural architecture is insufficient, the processor optimizes the next offer after the previous offer using a different trained neural architecture. The system according to claim 5, configured to execute the aforementioned instructions.

7. In response to the processor's determination that the accuracy of the trained neural architecture is satisfactory, it uses the trained neural architecture to optimize the next offer after the previous offer. The system according to claim 5, configured to execute the aforementioned instructions.

8. A method for selecting and using a neural architecture, Receiving negotiation traces, including offers from previous negotiations with the target negotiating party, The neural architecture is trained using only the received negotiation traces, Using the aforementioned trained neural architecture, optimize the offer to the target negotiating party during negotiations. Methods that include...

9. The process further includes generating the neural architecture based on the received architecture parameters, The received architecture parameters include architecture width and architecture depth, and training the neural architecture includes training the neural architecture generated based on the received architecture parameters. The method according to claim 8.

10. The method according to claim 8 or 9, further comprising training a plurality of neural architectures, wherein the plurality of neural architectures comprises the neural architectures.

11. Based on the simulated negotiation among the aforementioned multiple neural architectures, the neural architecture is selected as the optimal neural network from among the aforementioned multiple neural architectures. The method according to claim 10, further comprising:

12. During the aforementioned negotiations, a determination is made as to whether the accuracy of the trained neural architecture is satisfactory. The method according to any one of claims 8 to 11, further comprising:

13. In response to the determination that the accuracy of the aforementioned trained neural architecture is insufficient, optimize the next offer after the previous offer using a different trained neural architecture. The method according to claim 12, further comprising:

14. In response to the determination that the accuracy of the trained neural architecture is satisfactory, the trained neural architecture is used to optimize the next offer after the previous offer. The method according to claim 12, further comprising:

15. A non-temporary computer-readable medium configured to store instructions for selecting and using a neural architecture, The instruction causes the processor to perform an action, and the action is, Receiving negotiation traces, including offers from previous negotiations with the target negotiating party, The neural architecture is trained using only the received negotiation traces, Using the aforementioned trained neural architecture, optimize the offer to the target negotiating party during negotiations. Non-temporary computer-readable media, including [specific examples of such media].

16. The aforementioned operation involves generating a neural architecture based on the received architecture parameters. The present invention further includes the following: the received architecture parameters include architecture width and architecture depth, and training the neural architecture includes training the neural architecture generated based on the received architecture parameters. The non-temporary computer-readable medium according to claim 15.

17. The operation described above involves training multiple neural architectures, wherein the training includes the neural architecture described above. Based on the negotiation simulated among the aforementioned multiple neural architectures, the neural architecture is selected as the optimal neural network from among the aforementioned multiple neural architectures. A non-temporary computer-readable medium according to claim 15 or 16, further comprising:

18. The operation described above determines whether the accuracy of the trained neural architecture is satisfactory during the negotiation. A non-temporary computer-readable medium according to any one of claims 15 to 17, further comprising:

19. The operation involves, in response to a decision that the accuracy of the trained neural architecture is insufficient, optimizing the next offer after the previous offer using a different trained neural architecture. A non-temporary computer-readable medium according to claim 18, further comprising:

20. The operation is to use the trained neural architecture to optimize the next offer after the previous offer, in response to a decision that the accuracy of the trained neural architecture is satisfactory. A non-temporary computer-readable medium according to claim 18, further comprising:

Images