Implementation list

[gorkamunoz]

Task list based on PS tree and other things to implement.

Architectures

2-layer ECM

• #6
• Another thing

etc

[gorkamunoz]

Hey, I will take care of Basic 2-layer PS, I opened #6 for that!

[JPazem]

Hi everyone! Here is a list of the functions we thought would be nice to have for a first release, and the models that could be implemented with them and can help test the code.

Task list based on PS tree and other things to implement.

Methods to code:

Architecture

ECM_constructor(base_object, [N_clips_per_layer], adjacency_matrix)

• Make_ECM( ) : builds the initial ECM with a defined architecture, define whether use neural networks or not.

• ECM_modification( ) : modifies the ECM

  • by adding clips with clip bisections or weighted average of multiple clips
  • by adding layers
  • by adding edges

• Inductive_bias(N_excitations) : implement MePS’ inductive bias

• Tagging( ) : add some short-term memory to some transitions (eg. emoticons to avoid sampling bad rewards or sample only previously rewarded transitions)

Deliberate

• Make_probabilities(envelop_function): calculates the probabilities (given an envelop functions passed as an argument and to apply on the h-values);
• Deliberate(): takes a random walk by sampling. Function of tagging, inductive bias, and make_ECM();
• Reflect( ): implements the reflection mechanism.

Update

• update_reward_susceptibilities(): implement standard, ever or one-time increasing glow mechanism, or confidence mechanism
• update_h_values(): keywords include the method used for the update (standard, loss function minimization)

Others

• Pre-processor() : does the mapping from values on sensors to percepts
• Post-processor()

Models:

• Two-layer PS, with standard update rule
• Standard PS, including Multi-agent scenarios
• MePS
• FEPS
• DeepPS with neural network ECM and Graph NN

[gorkamunoz]

Hi @JPazem, could you maybe check from the previous list, what of the requested changes where already discussed during the last meeting? We recently merge (PR #8) all changes discussed during that meeting.

I will start working on the deep PS.

We also need to discuss about environments at some points...