Dip_ToM

TODO

Environment

Dipblue Agent

• Sensitive response of reject nego(for Messy ver)
• If counterpart attack me, reflect in relations(Messy ver)
• Check Retreat Phase of Order Handler

Data Preprocssing

Data preprocessing procedure are 3 stages. log_collector.py, one_epi, and Make Whole Data.

• target_dst.npy : shape : (num_data, 1) - where src go

• target_order.npy : shape : (num_data, 1) - what src do

• me_index.npy : shape : (num_data, num_past + 1) - index of power of me

• other_index.npy : shape : (num_data, num_past + 1) - index of power of other

• past_internal.npy : shape : (num_data, num_past, num_step, 6)

  • peaces[other], wars[other], trust_ratio[other] of me : 3
  • peaces[me], wars[me], trust_ratio[me] of other : 3
  • TODO -> now it is not like this, plz just use last 3 dim

• past_order.npy : shape : (num_data, num_past, num_step, 181)

  • order type [H, -, S, C] : 4
  • srcs location : 81
  • dsts location : 81
  • src_power : 7
  • dst_power : 7

• past_map_tactician.npy : shape : (num_data, num_past, num_step, 120)

  • Scalar map values of all regions (unit_type, loc) : 120
  • They are calculated from map_tactician adviser

• curr_internal.npy : shape : (num_data, curr_step, 6)

  • same as past
  • TODO -> now it is not like this, plz just use last 3 dim

• curr_map_tactician.npy : shape : (num_data, curr_step, 120)

  • same as past

• curr_src.npy : shape : (num_data, 81)

• me_weights.npy : shape : (num_data, 4)

  • TODO -> now it is not like this

• other_weights.npy : shape : (num_data, 4)

  • TODO -> now it is not like this

Codes

dipbluebot

DipBlue bot, DAIDE engine.
An implementation of "Dipblue: A diplomacy agent with strategic and trust reasoning"
We refer to Ferreira, André's code.

collect data

python utils/data_collector_new.py -na 7 -e 1000 -ev 100 -np 5 -ns 200 -n 0 --base_dir /data/cilab_ma/data/dip_tom --num_cpu 40

--num_agent, -na, : The number of players(dipblue) in a game.
--num_episode, -e : The number of train episode produced from a single cpu.
--num_eval, -ev :  The number of validation episode produced from a single cpu.
--number, -n : A run id for identification
--num_past, -np : The number of past transition for training in an episode. 
--num_step, -ns : The number of step 
--alpha, -a : 
--base_dir, -b : A directory for saved data.
--curr_step, -c : 
--data_type, -w : 
--num_cpu, -p : The number of cpu to collect data

run

python test.py