Processor for implementing reinforcement learning operations

A method for implementing predetermined reinforcement learning operations, said method comprising following computer-implemented steps:

• creating at least one reinforcement learning agent and at least one corresponding reinforcement learning environment using a first processor, and assigning a reinforcement learning agent ID to said reinforcement learning agent, and assigning a reinforcement learning environment ID to said reinforcement learning environment;
• creating an application-domain specific instruction set (ASI), said application-domain specific instruction set comprising instructions for implementing said predetermined reinforcement learning operations;
• creating each of said instructions by incorporating at least one of the reinforcement learning agent ID and the reinforcement learning environment ID therein, as an operand;
• storing said application-domain specific instruction set in a first memory module;
• selectively fetching at least one of said instructions, using a complex instruction fetch and decode (CISFD) unit, and decoding said instruction, and generating a plurality of predetermined threads corresponding to decoded instruction;
• embedding into the predetermined threads, at least one of said reinforcement learning agent ID and reinforcement learning environment ID associated with the decoded instruction;
• executing each of said predetermined threads in parallel using a second processor, and applying the corresponding instruction onto at least one of a reinforcement learning agent and reinforcement learning environment identified respectively by said reinforcement learning agent ID and reinforcement learning environment ID embedded in said predetermined threads, and determining at least one action, at least one state-value function, at least one Q-value and at least one reward value corresponding to the reinforcement learning agent, based on execution of the instruction corresponding to said predetermined threads;
• storing said action, state-value function, Q-value and reward value in predetermined memory partitions of a second memory module;
• accessing a neural network via a corresponding neural network data path, and training said neural network to perform at least one of a value-function approximation and a reward function approximation, based on said action, state-value function, Q-value and reward value.

A computer readable non-transitory storage medium having computer readable instructions stored thereupon, said computer readable instructions when executed by a computer processor, cause the computer processor to:

• create at least one reinforcement learning agent and at least one corresponding reinforcement learning environment, and assign a reinforcement learning agent ID to said reinforcement learning agent, and assign a reinforcement learning environment ID to said reinforcement learning environment;
• create an application-domain specific instruction set (ASI), said application-domain specific instruction set comprising instructions for implementing said predetermined reinforcement learning operations;
• create each of said instructions by incorporating at least one of said reinforcement learning agent ID and said reinforcement learning environment ID therein, as an operand;
• configure each of said instructions to be simultaneously implemented on a plurality of reinforcement learning agents and a plurality of reinforcement learning environments;
• selectively fetch at least one of said instructions by using a complex instruction fetch and decode (CISFD) unit, and decode said instruction, and generate a plurality of predetermined threads corresponding to decoded instruction;
• embed into the predetermined threads, at least one of said reinforcement learning agent ID and reinforcement learning environment ID associated with corresponding instruction;
• process each of said predetermined threads in parallel, and apply the corresponding instruction onto at least one of a reinforcement learning agent and reinforcement learning environment identified by the reinforcement learning agent ID and reinforcement learning environment ID embedded in said predetermined threads;
• determine at least one action, at least one state-value function, at least one Q-value and at least one reward value corresponding to the reinforcement learning agent, based on execution of the instruction corresponding to said predetermined threads;
• store said action, state-value function, Q-value and reward value in predetermined memory partitions of a second memory module;
• and
• access a neural network via a corresponding neural network data path, and train said neural network to perform at least one of value-function approximation and reward function approximation, based on said action, state-value function, Q-value and reward value.