add MADDPG algorithm #444
add MADDPG algorithm #444
Conversation
|
You can simply add those words after ReinforcementLearning.jl/.cspell/cspell.json Line 123 in 4973762 cc @pilgrimygy |
|
Thanks! And ask for suggestions about the implementation and code mistakes/style. |
|
I'll review it later tonight 😃 |
| function (π::MADDPGManager)(::PreEpisodeStage, ::AbstractEnv) | ||
| for (_, agent) in π.agents | ||
| if length(agent.trajectory) > 0 | ||
| pop!(agent.trajectory[:state]) | ||
| pop!(agent.trajectory[:action]) | ||
| if haskey(agent.trajectory, :legal_actions_mask) | ||
| pop!(agent.trajectory[:legal_actions_mask]) | ||
| end | ||
| end | ||
| end | ||
| end | ||
|
|
||
| function (π::MADDPGManager)(::PreActStage, env::AbstractEnv, actions) | ||
| # update each agent's trajectory | ||
| for (player, agent) in π.agents | ||
| push!(agent.trajectory[:state], state(env, player)) | ||
| push!(agent.trajectory[:action], actions[player]) | ||
| if haskey(agent.trajectory, :legal_actions_mask) | ||
| lasm = legal_action_space_mask(env, player) | ||
| push!(agent.trajectory[:legal_actions_mask], lasm) | ||
| end | ||
| end | ||
|
|
||
| # update policy | ||
| update!(π) | ||
| end | ||
|
|
||
| function (π::MADDPGManager)(::PostActStage, env::AbstractEnv) | ||
| for (player, agent) in π.agents | ||
| push!(agent.trajectory[:reward], reward(env, player)) | ||
| push!(agent.trajectory[:terminal], is_terminated(env)) | ||
| end | ||
| end | ||
|
|
||
| function (π::MADDPGManager)(::PostEpisodeStage, env::AbstractEnv) | ||
| # collect state and dummy action to each agent's trajectory | ||
| for (player, agent) in π.agents | ||
| push!(agent.trajectory[:state], state(env, player)) | ||
| push!(agent.trajectory[:action], rand(action_space(env))) | ||
| if haskey(agent.trajectory, :legal_actions_mask) | ||
| lasm = legal_action_space_mask(env, player) | ||
| push!(agent.trajectory[:legal_actions_mask], lasm) | ||
| end | ||
| end | ||
|
|
||
| # update policy | ||
| update!(π) | ||
| end |
There was a problem hiding this comment.
How about dispatching to the inner agent's corresponding methods?
Like calling agent(stage, env, action) in the for loop.
There was a problem hiding this comment.
Can you take a look at the NamedPolicy and see whether we can reuse existing code as much as possible? See also the MultiAgentManager
| temp_player = rand(keys(π.agents)) | ||
| t = π.agents[temp_player].trajectory |
| temp_player = rand(keys(π.agents)) | ||
| t = π.agents[temp_player].trajectory | ||
| inds = rand(π.rng, 1:length(t), π.batch_size) | ||
| batches = Dict((player, RLCore.fetch!(BatchSampler{SARTS}(π.batch_size), agent.trajectory, inds)) |
There was a problem hiding this comment.
The hardcoded SARTS will make the algorithm work only on environments of MINIMAL_ACTION_SET.
| s = vcat((batches[player][1] for (player, _) in π.agents)...) | ||
| a = vcat((batches[player][2] for (player, _) in π.agents)...) | ||
| s′ = vcat((batches[player][5] for (player, _) in π.agents)...) |
There was a problem hiding this comment.
vcat is not very efficient here. Try Flux.batch?
| s, a, s′ = send_to_host((s, a, s′)) | ||
| mu_actions = send_to_host(mu_actions) | ||
| new_actions = send_to_host(new_actions) |
There was a problem hiding this comment.
Thanks for your kind reviews! I'll check and update my codes later today.
|
Here is still a simple version of |
2 participants
PR Checklist
The description of the implementation is in discussion #404.
Here
MADDPGraises anunknown worderror... How can I fix it? @findmyway