Replicate-You
This game is from ldjam.com, which is currently unavailable.
You can try viewing the original page on Web Archive (may not load correctly).
You can try viewing the original page on Web Archive (may not load correctly).
Replicate-You: an experimental neural net based "programming by example" game. Since neural nets are a sort of modern equivalent of black magic, I think this is quite appropriate for the "summoning" theme.
Your goal is to summon a replica that plays a simple "pick up the coins and avoid the enemies" game as well as possible. To achieve this, you have to play the game yourself for some time, then you can summon a replica that copies the player's behaviour. Every time you let the replica play a level, the game gathers stats on its performance.
Use cursor keys or WSAD to move.
Note, you don't have to finish the levels to gather behavioural data: each move counts. What matters is that you make moves that optimise the behaviour of the replica. If you only play a particular level, you can expect the replica to be best at that level.
You can summon a working replica with as little as 50 moves, more for the higher levels. Plan your moves so the replica doesn't get stuck or hit too many enemies. If it doesn't work, try adding more moves and summon again.
Summoning will train a Tensorflow neural net that copies the behaviour of the player. Each time you move a step, a data point is created. Its input is a 9x9 map context area, its output is the move the player makes (left, right, up, down). The neural net consists of 2 convolution layers and 2 dense layers, and has only about 6,000 parameters, to keep training times tractable. Note that training is quite processor intensive, and requires GPU acceleration, and can make your browser stutter and lag. Training should take about 5-30 sec on a modern PC, depending on the size of the data set.
Your goal is to summon a replica that plays a simple "pick up the coins and avoid the enemies" game as well as possible. To achieve this, you have to play the game yourself for some time, then you can summon a replica that copies the player's behaviour. Every time you let the replica play a level, the game gathers stats on its performance.
Use cursor keys or WSAD to move.
Note, you don't have to finish the levels to gather behavioural data: each move counts. What matters is that you make moves that optimise the behaviour of the replica. If you only play a particular level, you can expect the replica to be best at that level.
You can summon a working replica with as little as 50 moves, more for the higher levels. Plan your moves so the replica doesn't get stuck or hit too many enemies. If it doesn't work, try adding more moves and summon again.
Summoning will train a Tensorflow neural net that copies the behaviour of the player. Each time you move a step, a data point is created. Its input is a 9x9 map context area, its output is the move the player makes (left, right, up, down). The neural net consists of 2 convolution layers and 2 dense layers, and has only about 6,000 parameters, to keep training times tractable. Note that training is quite processor intensive, and requires GPU acceleration, and can make your browser stutter and lag. Training should take about 5-30 sec on a modern PC, depending on the size of the data set.