|
| 1 | +Advanced Usage |
| 2 | +============== |
| 3 | +This section describes different optional user parameters available in DFO-LS. |
| 4 | + |
| 5 | +In the last section (:doc:`userguide`), we introduced :code:`dfols.solve()`, which has the optional input :code:`user_params`. This is a Python dictionary of user parameters. We will now go through the settings which can be changed in this way. More details are available in the paper [CFMR2018]_. |
| 6 | + |
| 7 | +The default values, used if no override is given, in some cases vary depending on whether :code:`objfun` has stochastic noise; that is, whether evaluating :code:`objfun(x)` several times at the same :code:`x` gives the same result or not. Whether or not this is the case is determined by the :code:`objfun_has_noise` input to :code:`dfols.solve()` (and not by inspecting :code:`objfun`, for instance). |
| 8 | + |
| 9 | +General Algorithm Parameters |
| 10 | +---------------------------- |
| 11 | +* :code:`general.rounding_error_constant` - Internally, all interpolation points are stored with respect to a base point :math:`x_b`; that is, we store :math:`\{y_t-x_b\}`, which reduces the risk of roundoff errors. We shift :math:`x_b` to :math:`x_k` when :math:`\|s_k\| \leq \text{const}\|x_k-x_b\|`, where 'const' is this parameter. Default is 0.1. |
| 12 | +* :code:`general.safety_step_thresh` - Threshold for when to call the safety step, :math:`\|s_k\| \leq \gamma_S \rho_k`. Default is :math:`\gamma_S =0.5`. |
| 13 | +* :code:`general.check_objfun_for_overflow` - Whether to cap the value of :math:`r_i(x)` when they are large enough that an OverflowError will be encountered when trying to evaluate :math:`f(x)`. Default is :code:`True`. |
| 14 | + |
| 15 | +Logging and Output |
| 16 | +------------------ |
| 17 | +* :code:`logging.n_to_print_whole_x_vector` - If printing all function evaluations to screen/log file, the maximum :code:`len(x)` for which the full vector :code:`x` should be printed also. Default is 6. |
| 18 | +* :code:`logging.save_diagnostic_info` - Flag so save diagnostic information at each iteration. Default is :code:`False`. |
| 19 | +* :code:`logging.save_poisedness` - If saving diagnostic information, whether to include the :math:`\Lambda`-poisedness of :math:`Y_k` in the diagnostic information. This is the most computationally expensive piece of diagnostic information. Default is :code:`True`. |
| 20 | +* :code:`logging.save_xk` - If saving diagnostic information, whether to include the full vector :math:`x_k`. Default is :code:`False`. |
| 21 | +* :code:`logging.save_rk` - If saving diagnostic information, whether to include the full vector :math:`[r_1(x_k)\:\cdots\:r_m(x_k)]`. The value :math:`f(x_k)` is always included. Default is :code:`False`. |
| 22 | + |
| 23 | +Initialization of Points |
| 24 | +------------------------ |
| 25 | +* :code:`init.random_initial_directions` - Build the initial interpolation set using random directions (as opposed to coordinate directions). Default is :code:`True`. |
| 26 | +* :code:`init.random_directions_make_orthogonal` - If building initial interpolation set with random directions, whether or not these should be orthogonalized. Default is :code:`True`. |
| 27 | +* :code:`init.run_in_parallel` - If using random directions, whether or not to ask for all :code:`objfun` to be evaluated at all points without any intermediate processing. Default is :code:`False`. |
| 28 | + |
| 29 | +Trust Region Management |
| 30 | +----------------------- |
| 31 | +* :code:`tr_radius.eta1` - Threshold for unsuccessful trust region iteration, :math:`\eta_1`. Default is 0.1. |
| 32 | +* :code:`tr_radius.eta2` - Threshold for very successful trust region iteration, :math:`\eta_2`. Default is 0.7. |
| 33 | +* :code:`tr_radius.gamma_dec` - Ratio to decrease :math:`\Delta_k` in unsuccessful iteration, :math:`\gamma_{dec}`. Default is 0.5 for smooth problems or 0.98 for noisy problems (i.e. :code:`objfun_has_noise = True`). |
| 34 | +* :code:`tr_radius.gamma_inc` - Ratio to increase :math:`\Delta_k` in very successful iterations, :math:`\gamma_{inc}`. Default is 2. |
| 35 | +* :code:`tr_radius.gamma_inc_overline` - Ratio of :math:`\|s_k\|` to increase :math:`\Delta_k` by in very successful iterations, :math:`\overline{\gamma}_{inc}`. Default is 4. |
| 36 | +* :code:`tr_radius.alpha1` - Ratio to decrease :math:`\rho_k` by when it is reduced, :math:`\alpha_1`. Default is 0.1 for smooth problems or 0.9 for noisy problems (i.e. :code:`objfun_has_noise = True`). |
| 37 | +* :code:`tr_radius.alpha2` - Ratio of :math:`\rho_k` to decrease :math:`\Delta_k` by when :math:`\rho_k` is reduced, :math:`\alpha_2`. Default is 0.5 for smooth problems or 0.95 for noisy problems (i.e. :code:`objfun_has_noise = True`). |
| 38 | + |
| 39 | +Termination on Small Objective Value |
| 40 | +------------------------------------ |
| 41 | +* :code:`model.abs_tol` - Tolerance on :math:`f(x_k)`; quit if :math:`f(x_k)` is below this value. Default is :math:`10^{-12}`. |
| 42 | +* :code:`model.rel_tol` - Relative tolerance on :math:`f(x_k)`; quit if :math:`f(x_k)/f(x_0)` is below this value. Default is :math:`10^{-20}`. |
| 43 | + |
| 44 | +Termination on Slow Progress |
| 45 | +---------------------------- |
| 46 | +* :code:`slow.history_for_slow` - History used to determine whether the current iteration is 'slow'. Default is 5. |
| 47 | +* :code:`slow.thresh_for_slow` - Threshold for objective decrease used to determine whether the current iteration is 'slow'. Default is :math:`10^{-4}`. |
| 48 | +* :code:`slow.max_slow_iters` - Number of consecutive slow successful iterations before termination (or restart). Default is :code:`20*len(x0)`. |
| 49 | + |
| 50 | +Stochastic Noise Information |
| 51 | +---------------------------- |
| 52 | +* :code:`noise.quit_on_noise_level` - Flag to quit (or restart) if all :math:`f(y_t)` are within noise level of :math:`f(x_k)`. Default is :code:`False` for smooth problems or :code:`True` for noisy problems. |
| 53 | +* :code:`noise.scale_factor_for_quit` - Factor of noise level to use in termination criterion. Default is 1. |
| 54 | +* :code:`noise.multiplicative_noise_level` - Multiplicative noise level in :math:`f`. Can only specify one of multiplicative or additive noise levels. Default is :code:`None`. |
| 55 | +* :code:`noise.additive_noise_level` - Additive noise level in :math:`f`. Can only specify one of multiplicative or additive noise levels. Default is :code:`None`. |
| 56 | + |
| 57 | +Interpolation Management |
| 58 | +-------------------------------- |
| 59 | +* :code:`interpolation.precondition` - whether or not to scale the interpolation linear system to improve conditioning. Default is :code:`True`. |
| 60 | + |
| 61 | +Regression Model Management |
| 62 | +--------------------------- |
| 63 | +* :code:`regression.num_extra_steps` - In successful iterations, the number of extra points (other than accepting the trust region step) to move, useful when :math:`|Y_k|>n+1` (:math:`n` is :code:`len(x0)`). Default is 0. |
| 64 | +* :code:`regression.increase_num_extra_steps_with_restart` - The amount to increase :code:`regression.num_extra_steps` by with each restarts, for instance if increasing the number of points with each restart. Default is 0. |
| 65 | +* :code:`regression.momentum_extra_steps` - If moving extra points in successful iterations, whether to use the 'momentum' method. If not, uses geometry-improving steps. Default is :code:`False`. |
| 66 | + |
| 67 | +Multiple Restarts |
| 68 | +----------------- |
| 69 | +* :code:`restarts.use_restarts` - Whether to do restarts when :math:`\rho_k` reaches :math:`\rho_{end}`, or (optionally) when all points are within noise level of :math:`f(x_k)`. Default is :code:`False` for smooth problems or :code:`True` for noisy problems. |
| 70 | +* :code:`restarts.max_unsuccessful_restarts` - Maximum number of consecutive unsuccessful restarts allowed (i.e.~restarts which did not reduce the objective further). Default is 10. |
| 71 | +* :code:`restarts.rhoend_scale` - Factor to reduce :math:`\rho_{end}` by with each restart. Default is 1. |
| 72 | +* :code:`restarts.use_soft_restarts` - Whether to use soft or hard restarts. Default is :code:`True`. |
| 73 | +* :code:`restarts.soft.num_geom_steps` - For soft restarts, the number of points to move. Default is 3. |
| 74 | +* :code:`restarts.soft.move_xk` - For soft restarts, whether to preserve :math:`x_k`, or move it to the best new point evaluated. Default is :code:`True`. |
| 75 | +* :code:`restarts.increase_npt` - Whether to increase :math:`|Y_k|` with each restart. Default is :code:`False`. |
| 76 | +* :code:`restarts.increase_npt_amt` - Amount to increase :math:`|Y_k|` by with each restart. Default is 1. |
| 77 | +* :code:`restarts.hard.increase_ndirs_initial_amt` - Amount to increase :code:`growing.ndirs_initial` by with each hard restart. To avoid a growing phase, it is best to set it to the same value as :code:`restarts.increase_npt_amt`. Default is 1. |
| 78 | +* :code:`restarts.hard.use_old_rk` - If using hard restarts, whether or not to recycle the objective value at the best iterate found when performing a restart. This saves one objective evaluation. Default is :code:`True`. |
| 79 | +* :code:`restarts.max_npt` - Maximum allowed value of :math:`|Y_k|`, useful if increasing with each restart. Default is :code:`npt`, the input parameter to :code:`dfols.solve()`. |
| 80 | +* :code:`restarts.soft.max_fake_successful_steps` - The maximum number of successful steps in a given run where the new (smaller) objective value is larger than the best value found in a previous run. Default is :code:`maxfun`, the input to :code:`dfols.solve()`. |
| 81 | +* :code:`restarts.auto_detect` - Whether or not to automatically determine when to restart. This is an extra condition, and restarts can still be triggered by small trust region radius, etc. Default is :code:`True`. |
| 82 | +* :code:`restarts.auto_detect.history` - How many iterations of data on model changes and trust region radii to store. There are two criteria used: trust region radius decreases (no increases over the history, more decreases than no changes), and change in model Jacobian (consistently increasing trend as measured by slope and correlation coefficient of line of best fit). Default is 30. |
| 83 | +* :code:`restarts.auto_detect.min_chgJ_slope` - Minimum rate of increase of :math:`\log(\|J_k-J_{k-1}\|_F)` over the past iterations to cause a restart. Default is 0.015. |
| 84 | +* :code:`restarts.auto_detect.min_correl` - Minimum correlation of the data set :math:`(k, \log(\|J_k-J_{k-1}\|_F))` required to cause a restart. Default is 0.1. |
| 85 | + |
| 86 | +Dynamically Growing Initial Set |
| 87 | +------------------------------- |
| 88 | +* :code:`growing.ndirs_initial` - Number of initial points to add (excluding :math:`x_k`). Default is :code:`npt-1`. |
| 89 | +* :code:`growing.num_new_dirns_each_iter` - Number of new search directions to add with each iteration where we do not have a full set of search directions. Default is 1. |
| 90 | +* :code:`growing.delta_scale_new_dirns` - When adding new search directions, the length of the step as a multiple of :math:`\Delta_k`. Default is 1, but if setting :code:`growing.perturb_trust_region_step=True` should be made smaller (e.g. 0.1). |
| 91 | +* :code:`growing.do_geom_steps` - While still growing the initial set, whether to do geometry-improving steps in the trust region algorithm, as per the usual algorithm. Default is :code:`False`. |
| 92 | +* :code:`growing.safety.do_safety_step` - While still growing the initial set, whether to perform safety steps, or the regular trust region steps. Default is :code:`True`. |
| 93 | +* :code:`growing.safety.reduce_delta` - While still growing the initial set, whether to reduce :math:`\Delta_k` in safety steps. Default is :code:`False`. |
| 94 | +* :code:`growing.safety.full_geom_step` - While still growing the initial set, whether to do a full geometry-improving step within safety steps (the same as the post-growing phase of the algorithm). Since this involves reducing :math:`\Delta_k`, cannot be :code:`True` if :code:`growing.safety.reduce_delta` is :code:`True`. Default is :code:`False`. |
| 95 | +* :code:`growing.full_rank.use_full_rank_interp` - Whether to perturb the interpolated :math:`J_k` to make it full rank, allowing the trust region step to include components in the full search space. If :code:`True`, setting :code:`growing.num_new_dirns_each_iter` to 0 is recommended. Default is :code:`False`. |
| 96 | +* :code:`growing.full_rank.scale_factor` - Magnitude of extra components added to :math:`J_k`. Default is :math:`10^{-2}`. |
| 97 | +* :code:`growing.full_rank.svd_scale_factor` - Floor singular values of :math:`J_k` at this factor of the last nonzero value. Default is 1. |
| 98 | +* :code:`growing.full_rank.min_sing_val` - Absolute floor on singular values of :math:`J_k`. Default is :math:`10^{-6}`. |
| 99 | +* :code:`growing.full_rank.svd_max_jac_cond` - Cap on condition number of :math:`J_k` after applying floors to singular values (effectively another floor on the smallest singular value, since the largest singular value is fixed). Default is :math:`10^8`. |
| 100 | +* :code:`growing.reset_delta` - Whether or not to reset trust region radius :math:`\Delta_k` to its initial value at the end of the growing phase. Default is :code:`False`. |
| 101 | +* :code:`growing.reset_rho` - Whether or not to reset trust region radius lower bound :math:`\rho_k` to its initial value at the end of the growing phase. Default is :code:`False`. |
| 102 | +* :code:`growing.gamma_dec` - Trust region decrease parameter during the growing phase. Default is :code:`tr_radius.gamma_dec`. |
| 103 | +* :code:`growing.perturb_trust_region_step` - Whether to perturb the trust region step by an orthogonal direction not yet searched. This is an alternative to :code:`growing.full_rank.use_full_rank_interp`. Default is :code:`False`. |
| 104 | + |
| 105 | + |
| 106 | +References |
| 107 | +---------- |
| 108 | + |
| 109 | +.. [CFMR2018] |
| 110 | + C. Cartis, J. Fiala, B. Marteau and L. Roberts, `Improving the Flexibility and Robustness of Model-Based Derivative-Free Optimization Solvers <https://arxiv.org/abs/1804.00154>`_, technical report, University of Oxford, (2018). |
| 111 | +
|
0 commit comments