| Set initial memory depth ; ; |
| All nodes and connections weight do not exist initially in STAMN; |
| The transition instance ; |
| A STAMN is constructed incrementally through interaction with the environment by the agent, |
| expanding the SATAMN until the transition prediction is contradicted with the current minimal |
| hypothesis model, reconstruct the hypothesis model by increase the memory depth . in this |
| paper, the constructing process includes the identifying and recalling simultaneously. |
| While one pattern or can be activated from the pattern cognition layer do |
| |
| for all existed state nodes |
| compute the identifying activation value by executing Algorithm 12 |
| If exist the then |
| the looped state is identified, and the weight , is adjusted according to (4), (5) |
| else |
| a new state node will be created, set . and the weight , is adjusted according to (4), (5) |
| end if |
| end for |
| for all existed state nodes |
| compute the transition prediction criterion by executing Algorithm 13. |
| If then |
| for the previous winner state node of the state node |
| set memory depth until each are discriminated |
| end for |
| reconstruct the hypothesis model by the new memory depth according to Algorithm 11 |
| end if |
| end for |
| for all existed action nodes |
| compute the activation value according to (9). |
| if exist the then |
| the weight is adjusted according to (6). |
| else |
| a new action node will be created, set . and the weight is adjusted according to (6). |
| end if |
| end for |
| The agent obtains the new observation vectors by = trans() |
| |
| end While |
