using NLPModels
using SolverTools


function coordinate_search3(nlp :: AbstractNLPModel;
			    tol :: Real = 1e-12,
			    α :: Float64 = 1.0,
			    x :: AbstractVector = nlp.meta.x0,
			    max_time :: Float64  = 3.0,
			    max_eval :: Int = -1,
			    β :: Real = 2.0,
			    greedy :: Bool = true)

  k = 0
  el_time = 0.0
  start_time = time()
  tired = neval_obj(nlp) > max_eval >= 0 || el_time > max_time
  optimal = α < tol
  T = eltype(x)
  status =:unknown
  @info log_header([:iter, :f, :H],[Int, T, T], hdr_override=Dict(:f=>"f(x)", :H=>"α"))

  f = obj(nlp,x)
  best_f = f
  best_i = 0

  while !(optimal || tired)
    @info log_row(Any[k, f, α])
    xt = copy(x)
    success = false
    done = false
    i = 1
    while !(done || i > nlp.meta.nvar)
      for s in [-1,1]
	xt[i] = x[i] + α * s
	f_xt = obj(nlp,xt)
	if (f_xt < best_f)
	  success = true
	  best_f = f_xt
	  best_i = i * s
	  if greedy
	    done = true
	    break
	  end
	end

	xt[i] = x[i]
      end
      i += 1
    end

    if success
      x[abs(best_i)] = x[abs(best_i)] + sign(best_i) *  α
      f = obj(nlp,x)
    else
      α /= β
    end

    k += 1
    el_time = time() - start_time
    tired = neval_obj(nlp) > max_eval >= 0 || el_time > max_time
    optimal = α <= tol
  end

  if (optimal)
    status =:unknown
  elseif (tired)
    if (neval_obj(nlp) > max_eval && max_eval >= 0)
      status =:max_eval
    elseif (el_time >= max_time)
      status =:max_time
    end
  end

  return GenericExecutionStats(status, nlp, solution=x, objective=f,
                                iter = k, elapsed_time = el_time)

end