This is the problem as far as I can see it:

• Each field has a random, unknown “danger level” ℓ_i that is evenly distributed across a certain interval [ℓ_min, ℓ_max]. So on average, the level is ℓ = (ℓ_min + ℓ_max)/2.
• Moving to the next field has a fixed cost Δc_m.
• Each move increases the “danger” d by Δd.

• If d > ℓ_i after a move, then two options can be taken:

  1. The total cost may increase by Δc_r = 2 (running away), or
  2. The total cost may increase by Δc_f = 15 (fighting), in which case d is reset to zero.

  The probability p(F) that d > ℓ_i, i.e. that a fight occurs, is 0 if d < ℓ_min, else (d – ℓ_min)/(ℓ_max – ℓ_min).

So our problem has the fixed parameters ℓ_min, ℓ_max, Δd, Δc_m, Δc_r, Δc_f and the state variables c and d. We have the following possible state transitions:

• No fight: With a probability of 1-p(F), nothing happens. then:

  c ← c + Δc_m
  d ← d + Δd

• A fight: With a probablity of p(F), we can choose the options

  1. c ← c + Δc_m + Δc_r
     d ← d + Δd
  2. c ← c + Δc_m + Δc_f
     d ← 0

Your questions seems to be how the total cost c can be minimized. In this statistical model, there is no definitive c. Instead, we will have to minimize the expected value for c. We can brute-force this by examining a decision tree. Each tree has can be assigned an expected cost, which can be calculated from the subtrees like

c = min(c_{F, f}, c_{F, r}) · p(F) + c_¬F · (1-p(F))

where c_{F, f} is the expected cost when fighting a fight at these conditions etc..

The following pseudocode gets the expected cost given a certain state:

function expected_cost(i, c, d) {
  return c if i == 0;

  p = probability_of_fight(d);

  c_no_fight      = (1-p) * expected_cost(i - 1, c + Δc_m,        d + Δd);

  c_fight_run, c_fight_fight = 0, 0;
  if (p > 0) {
    c_fight_run   =    p  * expected_cost(i - 1, c + Δc_m + Δc_r, d + Δd);
    c_fight_fight =    p  * expected_cost(i - 1, c + Δc_m + Δc_f, 0     );
  }

  return c_no_fight + min(c_fight_run, c_fight_fight);
}

Now all that has to be added is keeping track of which decisions were made in the winning paths – note that there are multiple winning paths, because either a fight may occur, or may not at each field. The decision we keep track of is whether the fight is carried out, or fled. Interesting analyses might be how often a fight is fled or fought – in this case, returning further values with the respective counts seems like the thing to do. Otherwise, inspecting a tree with the decision histories might be interesting.

Care should be taken to run that with very small i, as the tree can get very large (up to 3ⁱ). Memoization will help prevent some unneccessary calculations.

Depending on how you code the above algorithm, runtime can vary extremely. Below is a Perl script I used, which uses extensive caching to achieve a runtime in the order of seconds (if the current cost is a key in the cache as well, runtime is in the order of minutes). It performs an example analysis of counting how often running away and fighting is more beneficial. With those parameters, commiting to a fight isn’t a viable strategy before the game is 123 fields long.

use strict;
use warnings;
no warnings qw/recursion/;
use feature qw/state/;
use utf8;

use Math::BigInt;  # not neccessary. alternatively: no warnings 'imprecision'

# parameters as explained in answer
my ($ℓ_min, $ℓ_max, $Δd, $Δc_m, $Δc_r, $Δc_f) =
   (0xFF,   0xFF**2,113, 1,     2,     15   );

my ($steps) = @ARGV or die <<USAGE;
usage:
  $0 iterations

parameters:
  iterations -- the number of fields to traverse
USAGE

my ($cost, $fight, $run) = expected_cost($steps, 0);
printf "expected cost for i=%d steps is %gn",
    $steps, $cost;
printf "fight won %.3e timesn",
    $fight;
printf "run   won %.3e timesn",
    $run;
printf "fight:run relation: %sn",
    ($run > $fight) ? sprintf('1:%.3e', $run/$fight) : sprintf('%.3e:1', $fight/$run);


sub probability_of_fight {
    my ($d) = @_;
    return $d < $ℓ_min ? 0 : ($d - $ℓ_min)/($ℓ_max - $ℓ_min);
}

# Takes number $i of fields left and accumulated danger $d
# Return expected cost for that subtree, plus favourable fights and fleeings.
sub expected_cost {
    my ($i, $d) = @_;
    return 0, 0, 0 if $i == 0;

    state $cache = [];
    my $retval = $cache->[$i][$d] //= do {
        my $p = probability_of_fight($d);

        # variable prefixes:
        # nf -- no fight
        # f  -- fight
        # fr -- fight, but run away
        # ff -- fight, and carry out the fight

        my ($nf_c, $nf_fight, $nf_run)     = expected_cost($i - 1, $d + $Δd);

        my $f_c = 0;
        my ($f_fight, $f_run) = (Math::BigInt->new, Math::BigInt->new);
        if ($p > 0) {
            my ($fr_c, $fr_fight, $fr_run) = expected_cost($i - 1, $d + $Δd);
            $fr_c += $Δc_r;

            my ($ff_c, $ff_fight, $ff_run) = expected_cost($i - 1, 0       );
            $ff_c += $Δc_f;

            if ($ff_c < $fr_c) {
                # fighting is cheaper
                $f_c     = $ff_c;
                $f_fight = $ff_fight + 1;
                $f_run   = $ff_run;
            }
            elsif ($ff_c > $fr_c) {
                # running away is cheaper
                $f_c     = $fr_c;
                $f_fight = $fr_fight;
                $f_run   = $fr_run + 1;
            }
            else {
                # both are equivalent
                $f_c     = $ff_c;
                $f_fight = $ff_fight + $fr_fight;
                $f_run   = $ff_run   + $fr_run;
            }
        }

        my $expected_cost = $Δc_m + (1-$p) * $nf_c + $p * $f_c;
        [$expected_cost, $nf_fight + $f_fight, $nf_run + $f_run];
    };

    return @$retval;
}

Example session:

$ perl decide.pl 128
expected cost for i=128 steps is 154.27
fight won 2.432e+17 times
run   won 5.317e+36 times
fight:run relation: 1:2.186e+19

… where “won” means “occurs this often in the best decision tree”. This roughly translates to a probability of which decision is likely to be cheaper.

I had some time for further analysis. For each field, we can calculate if, when a fight occurs at this position, we should rather flee or fight. I plotted the result below:

… which uses the parameters from the above script.

How to read this graph: The x-axis is the distance from the end of the field, i.e. the number of steps (and thus fights) that can still occur. The y-axis is the danger level you have accumulated. If a player has to fight, he can find his distance-danger coordinate. If that coordinate is red, fighting will have the least cost. If the coordinate is colored blue, then fleeing is probably cheaper.

We can see that for example, during the last 7 steps fighting is never the best choice (the graph is already maxed out). Otherwise, the preference varies periodically. Fighting is most likely to be interesting at around 70–60 steps away from the end.