Robin Ranjit Singh Chauhan

Performance: allocating new nodes

7

Profiling using the scale testing branch (including the priority queue): php xdebug and kcachegrind A lot of time is spent constructing nodes, including __Constructor, and fromNode. Would you consider an...

Scale

8

Scale branch includes these changes: 1) Added large terrains to see performance on larger problems. 2) Replaced openList with SplPriorityQueue. Using hhvm on an old Ubuntu desktop, timings to find...

Maybe consider generalizing the algorithm to allow for BFS, UCS, DFS, etc

2

Generally the changes required to enable these are very small. But the algorithm would need to be written in a slightly more general way. Here is a nice example from...

Can I ask for comment about scalability of rule count and complexity? How would this compare to production rule algos like Rete? https://en.wikipedia.org/wiki/Rete_algorithm

Please make weights / checkpoint available

1

Otherwise we are all doing the same training each time, its wasteful. Thanks!

Goal directed

1

My understanding is Merlin targets "goal-directed behaviours". For example in the videos the agent repeatedly finds ways to a specific goal. However in the memory game, the cards are shuffled...

Simple agents

1

Thank you yosider for sharing this very interesting repo! Would you consider implementing simpler agents "RL-LSTM" and/or "RL-Mem" from the paper? They would make good baselines and are simpler to...