This repository contains the code used for the experiments in the paper Collaborative Decision Making Using Action Suggestions. You can find the paper here.
The development occurred using Julia v1.7 and v1.8. We recommend using the latest version of Julia.
First, clone the repo, change to the main folder, and run Julia.
git clone [email protected]:sisl/action_suggestions.git
cd action_suggestions
julia
We first need to activate the environment and include the supporting scripts. This process is scripted in setup.jl. You can run this file by:
julia> include("setup.jl")This repo contains polices and action value functions for RockSample(8, 4, 10, -1), Tag with the modified transition function, and the original implementation of Tag. You can start running those simulations immediately. Reference the Running Simulations section. To simulate the RockSample(7, 8, 20, 0) environment, you will need to generate the policy and action value matrix. Reference the Generating Policies section for directions on completing that process. The problems are referenced using the :rs84, :tag, and :tag_orig_tx Symbols. The RockSample(7, 8, 10, 0) problem has the :rs78 Symbol defined and ready for use after a policy and action value matrix is generated.
The simulation function is defined in run_sims.jl and is the run_sim function. See the doc string for detailed information about the arguments for this function. This file should be included when running the setup.jl script. However, if it was not, we can include this file by
julia> include("src/run_sims.jl")We can run a single simulation of the RockSample(8, 4, 10, -1) scenario by
julia> run_sim(:rs84)This command should produce an output similar to
Loading problem and policy...complete!
Agent: normal
Metric | Mean | Standard Dev | Standard Error | +/- 95 CI
--------------- | --------------- | --------------- | --------------- | ---------------
Reward | 11.07420 | NaN | NaN | NaN
Steps | 15.00000 | NaN | NaN | NaN
# Suggestions | 0.00000 | NaN | NaN | NaN
# Sugg / Step | 0.00000 | NaN | NaN | NaN
Details at each step can be output by setting the verbose keyword to true. A summary of key parameters will be output to the REPL at each time step. Recommend using verbose for single runs only (i.e. num_sims = 1)!
A visual depiction of the scenario can be output by setting the visualize keyword to true. Recommend using visualize for single runs only! In each scenario, a visualization of the belief is shown along with images before the suggestion is used to update the belief and after the suggestion is used to update the belief. The actions depicted on the bottom on in reference to the selected action with the displayed belief.
We can run multiple situations by using the num_sims keyword argument
julia> run_sim(:tag; num_sims=10)This command should produce an output similar to
Loading problem and policy...complete!
Running Simulations 100%|██████████████████████████████████████████████████| Time: 0:00:16 ( 1.61 s/it)
Agent: normal
Metric | Mean | Standard Dev | Standard Error | +/- 95 CI
--------------- | --------------- | --------------- | --------------- | ---------------
Reward | -10.08101 | 6.93397 | 2.19272 | 4.29772
Steps | 27.70000 | 16.22789 | 5.13171 | 10.05815
# Suggestions | 0.00000 | 0.00000 | 0.00000 | 0.00000
# Sugg / Step | 0.00000 | 0.00000 | 0.00000 | 0.00000
julia> run_sim(:rs84; num_sims=50, agent=:noisy, λ=1.0)
Loading problem and policy...complete!
Running Simulations 100%|██████████████████████████████████████████████████| Time: 0:00:00 ( 4.73 ms/it)
Agent: noisy, λ = 1.00
Metric | Mean | Standard Dev | Standard Error | +/- 95 CI
--------------- | --------------- | --------------- | --------------- | ---------------
Reward | 16.39763 | 3.55900 | 0.50332 | 0.98650
Steps | 17.44000 | 5.24564 | 0.74185 | 1.45402
# Suggestions | 5.42000 | 1.57907 | 0.22331 | 0.43770
# Sugg / Step | 0.32376 | 0.08556 | 0.01210 | 0.02372
julia> run_sim(:tag; num_sims=50, agent=:scaled, τ=0.75)
Loading problem and policy...complete!
Running Simulations 100%|██████████████████████████████████████████████████| Time: 0:00:31 ( 0.63 s/it)
Agent: scaled, τ = 0.75
Metric | Mean | Standard Dev | Standard Error | +/- 95 CI
--------------- | --------------- | --------------- | --------------- | ---------------
Reward | -1.61092 | 4.45021 | 0.62936 | 1.23354
Steps | 11.34000 | 6.36191 | 0.89971 | 1.76343
# Suggestions | 2.84000 | 2.15103 | 0.30420 | 0.59624
# Sugg / Step | 0.24639 | 0.10154 | 0.01436 | 0.02815
julia> run_sim(:rs84; num_sims=50, agent=:scaled, τ=0.75, msg_reception_rate=0.75)
Loading problem and policy...complete!
Running Simulations 100%|██████████████████████████████████████████████████| Time: 0:00:00 (11.81 ms/it)
Agent: scaled, τ = 0.75
Metric | Mean | Standard Dev | Standard Error | +/- 95 CI
--------------- | --------------- | --------------- | --------------- | ---------------
Reward | 15.58007 | 3.58591 | 0.50712 | 0.99396
Steps | 17.70000 | 6.02122 | 0.85153 | 1.66900
# Suggestions | 2.24000 | 0.79693 | 0.11270 | 0.22090
# Sugg / Step | 0.14602 | 0.09124 | 0.01290 | 0.02529
Function run_sim
Runs simulations and reports key metrics.
problem::Symbol: Problem to simulate (see RS_PROBS and TG_PROBS inconstants.jlfor options)
num_steps::Int=50: number of steps in each simulationnum_sims::Int=1: number of simulations to runverbose::Bool=false: print out details of each stepvisualize::Bool=false: render the environment at each step (2x per step)agent::Symbol=:normal: Which agent to simulate (see AGENTS for options)ν=1.0: hyperparameter for the naive agent (percent of suggestions to follow)τ=1.0: hyperparameter for the scaled agentλ=1.0: hyperparameter for the noisy agentmax_suggestions=Inf: Limit of the number of suggestions the agent can receivemsg_reception_rate=1.0: Reception rate of the agent for suggestionsperfect_v_random=1.0: Rate of perfect vs random suggestions (1.0=perfect, 0.0=random)init_rocks=nothing: For RockSamplePOMDP only. Designate the state of initial rocks. Must be a vector with length equal to the number of rocks (e.g. [1, 0, 0, 1])suggester_belief=[1.0, 0.0]: RockSamplePOMDP only. Designate the initial belief over good rocks and bad rocks respectively. [1.0, 0.0] = perfect knowledge suggester, [0.75, 0.5] would represent a suggester with a bit more knowledge over good rocks but no additional information for the bad rocks.init_pos=nothing: TagPOMDP only. Set the initial positions of the agent and opponent. The form is Vector{Tuple{Int, Int}}. E.g. [(1, 1), (5, 2)].rng=Random.GLOBAL_RNG: Provide a random number generator
The function to generate and save policies is in pol_generator.jl and the function to generate and save the action value function as a matrix is contained in generate_q.jl. Both of these files are included by the setup.jl script but can be included manually if needed.
To generate and save a policy, call generate_problem_and_policy with the problem of interest. Parameters can be passed to the SARSOP solver by keywords. For the RockSample(7, 8, 20, 0) results contained in the paper, a timeout value of 10800 was used.
julia> generate_problem_and_policy(:rs78; timeout=300)
Generating a pomdpx file: model.pomdpx
Loading the model ...
input file : model.pomdpx
loading time : 301.06s
SARSOP initializing ...
initialization time : 0.88s
-------------------------------------------------------------------------------
Time |#Trial |#Backup |LBound |UBound |Precision |#Alphas |#Beliefs
-------------------------------------------------------------------------------
0.88 0 0 7.35092 28.5048 21.1539 13 1
0.95 2 50 7.35092 27.1536 19.8027 10 24
1.01 5 101 11.7638 25.7925 14.0287 22 38
1.1 7 150 12.3727 25.6247 13.2519 52 63
1.22 9 203 12.3727 25.5254 13.1526 78 84
...
...
263.85 389 9057 15.3982 22.4768 7.07857 2242 3236
267.16 391 9100 15.3982 22.4738 7.0756 2285 3250
269.68 393 9157 15.3982 22.4723 7.07412 2342 3269
272.37 395 9213 15.3982 22.4701 7.07186 2398 3287
275.73 397 9259 15.3982 22.4665 7.06831 2271 3305
277.82 399 9301 15.3982 22.4628 7.06459 2313 3318
281.28 401 9350 15.3982 22.4527 7.05447 2362 3336
284.72 403 9400 15.3982 22.4422 7.04402 2412 3356
286.41 405 9455 15.3982 22.4306 7.03245 2467 3376
289.55 407 9500 15.3982 22.4241 7.0259 2340 3391
293.48 410 9550 15.3982 22.4148 7.0166 2390 3406
297.99 413 9607 15.3982 22.4066 7.00837 2447 3424
301.29 415 9657 15.3982 22.4044 7.00621 2497 3441
-------------------------------------------------------------------------------
SARSOP finishing ...
Preset timeout reached
Timeout : 300.000000s
Actual Time : 301.290000s
-------------------------------------------------------------------------------
Time |#Trial |#Backup |LBound |UBound |Precision |#Alphas |#Beliefs
-------------------------------------------------------------------------------
301.51 415 9657 15.3982 22.4044 7.00621 2356 3441
-------------------------------------------------------------------------------
Writing out policy ...
output file : policy.out
Complete! Saved as: policies/rs_7-8-20-0_pol.jld2
generate_and_save_Q(:rs78)
Loading problem and policy...complete!
Calculating action value matrix 100%|████████████████████████████████████████| Time: 1:13:19 (23.60 ms/it)
After generating the policy and the Q matrix, you should have two more files saved in the policies folder (rs_7-8-20-0_pol.jld2 and rs_7-8-20-0_Q.jld2). Now you can run simulations with the :rs78 symbol as described in the Running Simulations section.
@inproceedings{Asmar2022},
title = {Collaborative Decision Making Using Action Suggestions},
author = {Dylan M. Asmar and Mykel J. Kochenderfer},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2022}