Autonomous supply chain by collaborative software agents and reinforcement learning
A system and method are disclosed to train machine learning models, generate software agents, and evaluate, via reinforcement learning, the actions of the software agents in a simulated ecosystem. Embodiments include a computer comprising a processor and memory and configured to train one or more machine learning models to generate one or more software agents, wherein each software agent comprises an autonomous software program designed to execute a task in a supply chain network. Embodiments generate a first software agent and a second software agent, and a simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks. Embodiments simulate one or more tasks executed by the software agents in the simulated supply chain ecosystem, review the tasks according to one or more defined objectives, and apply reinforcement incentives to the software agents.
1. A computer-implemented method, comprising:
training, by a computer comprising a processor and memory, one or more machine learning models of a Markov-based reinforcement learning process;
generating, with the one or more machine learning models, a first software agent and a second software agent, wherein each of the one or more software agents comprises an autonomous software program designed to execute one or more tasks in a supply chain network;
generating, by the computer, a simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulating, by the computer, the one or more tasks executed by the first software agent and the second software agent in the simulated supply chain ecosystem, wherein decisions by the first software agent and the second software agent to execute the one or more tasks in the simulated supply chain ecosystem are made according to the Markov-based reinforcement learning process and at least partially based on an epsilon-greedy approach;
reviewing, by the computer, the one or more tasks executed by the first software agent and the second software agent according to one or more defined objectives; and
applying, by the computer, reinforcement incentives to the first software agent and the second software agent, based at least in part on achievement of the one or more defined objectives.
2. The computer-implemented method of claim 1 , further comprising, after applying reinforcement incentives to the first software agent and the second software agent:
generating, by the computer, an updated simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulating, by the computer, one or more updated tasks executed by the first software agent and the second software agent in the updated simulated supply chain ecosystem;
reviewing, by the computer, the one or more updated tasks executed by the first software agent and the second software agent according to the one or more defined objectives; and
applying, by the computer, updated reinforcement incentives to the first software agent and the second software agent, based at least in part on the degree to which the updated tasks executed by the first software agent and the second software agent achieved the one or more defined objectives.
3. The computer-implemented method of claim 1 , further comprising:
generating, with the one or more machine learning models, a first software agent designed to execute a task in the supply chain network by selecting and executing one or more actions associated with a highest available Q-value, wherein the Q-value represents a sum of an immediate reward and delayed future rewards of the associated action.
4. The computer-implemented method of claim 1 , further comprising the computer generating the reinforcement incentives using one or more policy-gradient methods, wherein the one or more policy-gradient methods comprise, at least in part, a gradient of parameters generated by a neural network.
5. The computer-implemented method of claim 1 , wherein at least one of the first software agent and the second software agent comprises a software agent selected from a group consisting of:
a distribution planner software agent;
a route planner software agent;
a replenishment planner software agent;
a packaging and assembly planner software agent;
a production planner software agent;
a logistics planner software agent; and
an in-store category manager software agent.
6. The computer-implemented method of claim 1 , wherein the one or more machine learning models generates a first software agent and a second software agent using demand forecasts.
7. The computer-implemented method of claim 1 , wherein the simulated supply chain ecosystem comprises an end-to-end simulated ecosystem in which the first software agent and the second software agent interact with one another while executing tasks.
8. A system comprising a computer, the computer comprising a processor and memory and configured to:
train one or more machine learning models of a Markov-based reinforcement learning process;
generate, with the one or more machine learning models, a first software agent and a second software agent, wherein each of the one or more software agents comprises an autonomous software program designed to execute one or more tasks in a supply chain network;
generate a simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulate the one or more tasks executed by the first software agent and the second software agent in the simulated supply chain ecosystem, wherein decisions by the first software agent and the second software agent to execute the one or more tasks in the simulated supply chain ecosystem are made according to the Markov-based reinforcement learning process and at least partially based on an epsilon-greedy approach;
review the one or more tasks executed by the first software agent and the second software agent according to one or more defined objectives; and
apply reinforcement incentives to the first software agent and the second software agent, based at least in part on achievement of the one or more defined objectives.
9. The system of claim 8 , further comprising the computer, after applying reinforcement incentives to the first software agent and the second software agent:
generating an updated simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulating one or more updated tasks executed by the first software agent and the second software agent in the updated simulated supply chain ecosystem;
reviewing the one or more updated tasks executed by the first software agent and the second software agent according to the one or more defined objectives; and
applying updated reinforcement incentives to the first software agent and the second software agent, based at least in part on the degree to which the updated tasks executed by the first software agent and the second software agent achieved the one or more defined objectives.
10. The system of claim 8 , further comprising the computer:
generating, with the one or more machine learning models, a first software agent designed to execute a task in the supply chain network by selecting and executing one or more actions associated with a highest available Q-value, wherein the Q-value represents a sum of an immediate reward and delayed future rewards of the associated action.
11. The system of claim 8 , further comprising the computer generating the reinforcement incentives using one or more policy-gradient methods, wherein the one or more policy-gradient methods comprise, at least in part, a gradient of parameters generated by a neural network.
12. The system of claim 8 , wherein at least one of the first software agent and the second software agent comprises a software agent selected from a group consisting of:
a distribution planner software agent;
a route planner software agent;
a replenishment planner software agent;
a packaging and assembly planner software agent;
a production planner software agent;
a logistics planner software agent; and
an in-store category manager software agent.
13. The system of claim 8 , wherein the one or more machine learning models generates a first software agent and a second software agent using demand forecasts.
14. The system of claim 8 , wherein the simulated supply chain ecosystem comprises an end-to-end simulated ecosystem in which the first software agent and the second software agent interact with one another while executing tasks.
15. A non-transitory computer-readable storage medium embodied with software, the software when executed configured to:
train one or more machine learning models of a Markov-based reinforcement learning process;
generate, with the one or more machine learning models, a first software agent and a second software agent, wherein each of the one or more software agents comprises an autonomous software program designed to execute one or more tasks in a supply chain network;
generate a simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulate the one or more tasks executed by the first software agent and the second software agent in the simulated supply chain ecosystem, wherein decisions by the first software agent and the second software agent to execute the one or more tasks in the simulated supply chain ecosystem are made according to the Markov-based reinforcement learning process and at least partially based on an epsilon-greedy approach;
review the one or more tasks executed by the first software agent and the second software agent according to one or more defined objectives; and
apply reinforcement incentives to the first software agent and the second software agent, based at least in part on achievement of the one or more defined objectives.
16. The non-transitory computer-readable storage medium of claim 15 , wherein the software when executed is further configured to, after applying reinforcement incentives to the first software agent and the second software agent:
generate an updated simulated supply chain ecosystem representing a hierarchical structure of supply chain network tasks;
simulate one or more updated tasks executed by the first software agent and the second software agent in the updated simulated supply chain ecosystem;
review the one or more updated tasks executed by the first software agent and the second software agent according to the one or more defined objectives; and
apply updated reinforcement incentives to the first software agent and the second software agent, based at least in part on the degree to which the updated tasks executed by the first software agent and the second software agent achieved the one or more defined objectives.
17. The non-transitory computer-readable storage medium of claim 15 , wherein the software when executed is further configured to:
generate, with the one or more machine learning models, a first software agent designed to execute a task in the supply chain network by selecting and executing one or more actions associated with a highest available Q-value, wherein the Q-value represents a sum of an immediate reward and delayed future rewards of the associated action.
18. The non-transitory computer-readable storage medium of claim 15 , wherein the software when executed is further configured to:
generate the reinforcement incentives using one or more policy-gradient methods, wherein the one or more policy-gradient methods comprise, at least in part, a gradient of parameters generated by a neural network.
19. The non-transitory computer-readable storage medium of claim 15 , wherein at least one of the first software agent and the second software agent comprises a software agent selected from a group consisting of:
a distribution planner software agent;
a route planner software agent;
a replenishment planner software agent;
a packaging and assembly planner software agent;
a production planner software agent;
a logistics planner software agent; and
an in-store category manager software agent.
20. The non-transitory computer-readable storage medium of claim 15 , wherein the one or more machine learning models generates a first software agent and a second software agent using demand forecasts.