smbot/ars.lua

-- Augmented Random Search
-- https://arxiv.org/abs/1803.07055
-- with some tweaks (lips) by myself.

local abs = math.abs
local floor = math.floor
local ipairs = ipairs
local max = math.max
local print = print

local Base = require "Base"

local nn = require "nn"
local normal = nn.normal
local prod = nn.prod
local zeros = nn.zeros

local util = require "util"
local argsort = util.argsort
local calc_mean_dev = util.calc_mean_dev

local Ars = Base:extend()

local function collect_best_indices(scored, top, antithetic)
    -- select one (the best) reward of each pos/neg pair.
    local best_rewards
    if antithetic then
        best_rewards = {}
        for i = 1, #scored, 2 do
            local ind = floor(i / 2) + 1
            local pos = scored[i + 0]
            local neg = scored[i + 1]
            best_rewards[ind] = max(pos, neg)
        end
    else
        best_rewards = scored
    end

    local indices = argsort(best_rewards, function(a, b) return a > b end)

    for i = top + 1, #best_rewards do indices[i] = nil end
    return indices
end

local function kinda_lipschitz(dir, pos, neg, mid)
    local _, dev = calc_mean_dev(dir)
    local c0 = neg - mid
    local c1 = pos - mid
    local l0 = abs(3 * c1 + c0)
    local l1 = abs(c1 + 3 * c0)
    return max(l0, l1) / (2 * dev)
end

local function amsgrad(step) -- in-place! -- TODO: fix this.
    if mom1 == nil then mom1 = nn.zeros(#step) end
    if mom2 == nil then mom2 = nn.zeros(#step) end
    if mom2max == nil then mom2max = nn.zeros(#step) end

    local b1_t = pow(cfg.adam_b1, epoch_i)
    local b2_t = pow(cfg.adam_b2, epoch_i)

    -- NOTE: with LuaJIT, splitting this loop would
    --       almost certainly be faster.
    for i, v in ipairs(step) do
        mom1[i] = cfg.adam_b1 * mom1[i] + (1 - cfg.adam_b1) * v
        mom2[i] = cfg.adam_b2 * mom2[i] + (1 - cfg.adam_b2) * v * v
        mom2max[i] = max(mom2[i], mom2max[i])
        if cfg.adam_debias then
            local num = (mom1[i] / (1 - b1_t))
            local den = sqrt(mom2max[i] / (1 - b2_t)) + cfg.adam_eps
            step[i] = num / den
        else
            step[i] = mom1[i] / (sqrt(mom2max[i]) + cfg.adam_eps)
        end
    end
end

function Ars:init(dims, popsize, poptop, learning_rate, sigma, antithetic)
    self.dims = dims
    self.popsize = popsize or 4 + (3 * floor(log(dims)))
    self.learning_rate = learning_rate or 3/5 * (3 + log(dims)) / (dims * sqrt(dims))
    self.sigma = sigma or 1
    self.antithetic = antithetic and true or false

    self.poptop = poptop or popsize
    assert(self.poptop <= popsize)
    if self.antithetic then self.popsize = self.popsize * 2 end

    self._params = nn.zeros(self.dims)
end

function Ars:params(new_params)
    if new_params ~= nil then
        assert(#self._params == #new_params, "new parameters have the wrong size")
        for i, v in ipairs(new_params) do self._params[i] = v end
    end
    return self._params
end

function Ars:ask(graycode)
    local asked = {}
    local noise = {}

    for i = 1, self.popsize do
        local asking = zeros(self.dims)
        local noisy = zeros(self.dims)
        asked[i] = asking

        if self.antithetic and i % 2 == 0 then
            for j, v in ipairs(self._params) do
                asking[i] = v - noisy[j]
            end

        else
            if graycode ~= nil then
                for j = 1, self.dims do
                    noisy[j] = exp(-precision * nn.uniform())
                end
                for j = 1, self.dims do
                    noisy[j] = nn.uniform() < 0.5 and noisy[j] or -noisy[j]
                end
            else
                for j = 1, self.dims do
                    noisy[j] = self.sigma * nn.normal()
                end
            end

            for j, v in ipairs(self._params) do
                asking[j] = v + noisy[j]
            end
        end

        noise[i] = noisy
    end

    self.noise = noise
    return asked, noise
end

function Ars:tell(scored, unperturbed_score)
    local indices = collect_best_indices(scored, self.poptop, self.antithetic)
    --print("best trials:", indices)

    local top_rewards = {}
    for i = 1, #scored do top_rewards[i] = 0 end
    for _, ind in ipairs(indices) do
        local sind = (ind - 1) * 2 + 1
        top_rewards[sind + 0] = scored[sind + 0]
        top_rewards[sind + 1] = scored[sind + 1]
    end
    --print("top:", top_rewards)

    if self.antithetic then
        local top_delta_rewards = {} -- only used for printing.
        for i, ind in ipairs(indices) do
            local sind = (ind - 1) * 2 + 1
            top_delta_rewards[i] = abs(top_rewards[sind + 0] - top_rewards[sind + 1])
        end
        --print("best deltas:", top_delta_rewards)
    end

    local step = nn.zeros(self.dims)
    local _, reward_dev = calc_mean_dev(top_rewards)
    if reward_dev == 0 then reward_dev = 1 end

    if self.antithetic then
        for i = 1, floor(self.popsize / 2) do
            local ind = (i - 1) * 2 + 1
            local pos = top_rewards[ind + 0]
            local neg = top_rewards[ind + 1]
            local reward = pos - neg
            if reward ~= 0 then
                local noisy = self.noise[i]

                if unperturbed_score ~= nil then
                    local lips = kinda_lipschitz(noisy, pos, neg, unperturbed_score)
                    reward = reward / lips / self.sigma
                else
                    reward = reward / reward_dev
                end

                for j, v in ipairs(noisy) do
                    step[j] = step[j] + reward * v / self.poptop
                end
            end
        end
    else
        for i = 1, self.popsize do
            local reward = top_rewards[i] / reward_dev
            if reward ~= 0 then
                local noisy = self.noise[i]

                for j, v in ipairs(noisy) do
                    step[j] = step[j] + reward * v / self.poptop
                end
            end
        end
    end

    for i, v in ipairs(self._params) do
        self._params[i] = v + self.learning_rate * step[i]
    end

    self.asked = nil
end

return {
    Ars = Ars,
}